https://github.com/yhirose/mtlcpp
Utilities for Metal-cpp
https://github.com/yhirose/mtlcpp
apple-silicon cpp cpp20 header-only m1-mac metal metal-cpp
Last synced: 6 months ago
JSON representation
Utilities for Metal-cpp
- Host: GitHub
- URL: https://github.com/yhirose/mtlcpp
- Owner: yhirose
- License: mit
- Created: 2023-10-03T21:11:29.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2024-06-15T20:01:13.000Z (over 1 year ago)
- Last Synced: 2025-04-18T10:22:51.053Z (6 months ago)
- Topics: apple-silicon, cpp, cpp20, header-only, m1-mac, metal, metal-cpp
- Language: C++
- Homepage:
- Size: 1.98 MB
- Stars: 10
- Watchers: 5
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
mtlcpp
======
A header-only C++20 linear algebra library for Metal on MacOS
* This project is still in development and is far from reaching the first alpha version :)
* Data types supported in this library are `int` and `float` only, since Metal doesn't support `double`
* This library uses GPU cores in Apple M1 chip with [Metal-cpp](https://developer.apple.com/metal/cpp/)
Build and run unit tests and benchmark
--------------------------------------
* Install Xcode Command Line Tools
* Run the following commands in Terminal
```bash
cd test
make
```
Benchmark as of 3/2/2024 on M1 MacBook Pro 14
---------------------------------------------
| ns/op | op/s | benchmark
|--------------------:|--------------------:|:----------
| 150,856,709.00 | 6.63 | CPU: `a + b`
| 2,262,442.07 | 442.00 | GPU: `a + b`
| 1,351,401.59 | 739.97 | Eigen: `a + b`
| 964,220,500.00 | 1.04 | CPU: `a.dot(b)`
| 1,094,602.35 | 913.57 | GPU: `a.dot(b)`
| 3,002,299.36 | 333.08 | Eigen: `a * b`
```cpp
// test/bench.cpp
// `add` benchmark
const size_t n = 10'000'000;
auto a = mtl::ones({n});
auto b = mtl::ones({n});
auto c = mtl::array();
mtl::use_cpu();
Bench().run("CPU: a + b", [&] {
c = a + b;
});
mtl::use_gpu();
Bench().run("GPU: a + b", [&] {
c = a + b;
});
auto aa = Eigen::Vector::Ones(n);
auto bb = Eigen::Vector::Ones(n);
auto cc = Eigen::Vector(n);
Bench().run("Eigen: a + b", [&] {
cc = aa + bb;
});
// `dot` benchmark
auto a = mtl::ones({1000, 1000});
auto b = mtl::ones({1000, 100});
auto c = mtl::array();
mtl::use_cpu();
Bench().run("CPU: a.dot(b)", [&] {
c = a.dot(b);
});
mtl::use_gpu();
Bench().run("GPU: a.dot(b)", [&] {
c = a.dot(b);
});
auto aa = Eigen::Matrix::Ones(1000, 1000);
auto bb = Eigen::Matrix::Ones(1000, 100);
auto cc = Eigen::Matrix();
Bench().run("Eigen: a * b", [&] {
cc = aa * bb;
});
```
Operations
----------
### GPU and CPU
* `+` (add)
* `-` (sub)
* `*` (mul)
* `/` (div)
* `dot` (dot product)
### CPU only
* `==`
* `clone`
* `constants`
* `empty`
* `zeros`
* `ones`
* `random`
* `transpose`
* `sigmoid`
* `sum`
* `mean`
* `min`
* `max`
* `count`
* `all`
* `softmax`
* `argmax`
* `array_equal`
* `allclose`
License
-------
MIT license (© 2024 Yuji Hirose)