vectorized-fields

Vectorized AVX-512 256-bit arithmetic library focused on BN254 field vector operations with Rust bindings. Boasts 1-3x speedups over plain CPU arithmetic.

The library performs elementwise operations on vectors:

$$ \mathbf{z} = \mathbf{x} \odot \mathbf{y} = \begin{bmatrix} x_0 \cdot y_0 \\ x_1 \cdot y_1 \\ x_2 \cdot y_2 \\ x_3 \cdot y_3 \end{bmatrix} $$

$$ \mathbf{z} = \mathbf{x} + \mathbf{y} = \begin{bmatrix} x_0 + y_0 \\ x_1 + y_1 \\ x_2 + y_2 \\ x_3 + y_3 \end{bmatrix} $$

$$ \mathbf{z} = \mathbf{x} - \mathbf{y} = \begin{bmatrix} x_0 - y_0 \\ x_1 - y_1 \\ x_2 - y_2 \\ x_3 - y_3 \end{bmatrix} $$

$$ z = \mathbf{x} \cdot \mathbf{y} = \begin{bmatrix} x_0 \\ x_1 \\ x_2 \\ x_3 \end{bmatrix} \cdot \begin{bmatrix} y_0 \\ y_1 \\ y_2 \\ y_3 \end{bmatrix} = x_0 \cdot y_0 + x_1 \cdot y_1 + x_2 \cdot y_2 + x_3 \cdot y_3 $$

Features

This library provides optimized vector operations for BN254 field arithmetic, leveraging AVX-512 instructions. The following operations are supported:

• Vector operations: add, sub, mul, sum, inner_product
• Parallel versions: add_par, sub_par, mul_par, sum_par, inner_product_par
• In-place operations: add_inplace, sub_inplace, mul_inplace

All operations are suffixed with _bn254.

These operations use AVX-512 instructions for efficiency.

Benchmarks

Note

Benchmarks performed on an AWS c7a.8xlarge instance.

sragss/gpu-sumcheck at 6971f52

Future

Please get in touch with @samrags_ if you'd like specific changes or are curious if this works with your workload.