4 releases (breaking)

0.4.0 Jan 14, 2025
0.3.0 Oct 28, 2024
0.2.0 Aug 27, 2024
0.1.1 Jul 19, 2024

#100 in Algorithms

Download history 1013/week @ 2024-12-04 1128/week @ 2024-12-11 553/week @ 2024-12-18 240/week @ 2024-12-25 578/week @ 2025-01-01 797/week @ 2025-01-08 1490/week @ 2025-01-15 1338/week @ 2025-01-22 1485/week @ 2025-01-29 1444/week @ 2025-02-05 1384/week @ 2025-02-12 4463/week @ 2025-02-19 3733/week @ 2025-02-26 3703/week @ 2025-03-05 3166/week @ 2025-03-12 2887/week @ 2025-03-19

14,290 downloads per month
Used in 32 crates (4 directly)

MIT/Apache

1MB
28K SLoC

CubeCL Linear Algebra Library.

The crate contains common linear algebra algorithms.

Algorithms

  • Tiling 2D Matrix Multiplication.

    The kernel is very flexible and can be used on pretty much any hardware.

  • Cooperative Matrix Multiplication.

    The kernel is using Automatic Mixed Precision (AMP) to leverage cooperative matrix-multiply and accumulate instructions. For f32 tensors, the inputs are casted into f16, but the accumulation is still performed in f32. This may cause a small lost in precision, but with way faster execution.

Benchmarks

You can run the benchmarks from the workspace with the following:

cargo bench --bench matmul --features wgpu # for wgpu
cargo bench --bench matmul --features cuda # for cuda

On an RTX 3070 we get the following results:

matmul-wgpu-f32-tiling2d

―――――――― Result ―――――――――
  Samples     100
  Mean        13.289ms
  Variance    28.000ns
  Median      13.271ms
  Min         12.582ms
  Max         13.768ms
―――――――――――――――――――――――――
matmul-cuda-f32-tiling2d

―――――――― Result ―――――――――
  Samples     100
  Mean        12.754ms
  Variance    93.000ns
  Median      12.647ms
  Min         12.393ms
  Max         14.501ms
―――――――――――――――――――――――――
matmul-cuda-f32-cmma

―――――――― Result ―――――――――
  Samples     100
  Mean        4.996ms
  Variance    35.000ns
  Median      5.084ms
  Min         4.304ms
  Max         5.155ms
―――――――――――――――――――――――――

Dependencies

~5–19MB
~199K SLoC