e3nn.c

Pure C implementation of e3nn. Mostly done for pedagogical reasons, but similar code could be used for C/C++ implementations of e3nn-based models for inference or CUDA kernels for faster operations within Python libraries.

Currently the only operations implemented are the tensor product, and spherical harmonics.

Single-thread CPU performance of the tensor product on an Intel i5 Desktop Processor.

Message Computation

#include <stdio.h>

#include "e3nn.h"

// example.c
int main(void){

    float node_position_sh[9] = { 0 };
    Irreps* node_irreps = irreps_create("1x0e + 1x1o + 1x2e");
    spherical_harmonics(node_irreps, 1, 2, 3, node_position_sh);

    printf("sh ["); for (int i = 0; i < 9; i++){ printf("%.2f, ", node_position_sh[i]); } printf("]\n");
    irreps_free(node_irreps);

    float neighbor_feature[] = { 7, 8, 9 };
    float product[27] = { 0 };
    Irreps* node_sh_irreps = irreps_create("1x0e + 1x1o + 1x2e");
    Irreps* neighbor_feature_irreps = irreps_create("1x1e");
    Irreps* product_irreps = irreps_create("1x0o + 1x1o + 2x1e + 1x2e + 1x2o + 1x3e");
    tensor_product(node_sh_irreps, node_position_sh, 
                   neighbor_feature_irreps, neighbor_feature, 
                   product_irreps, product);

    printf("product ["); for (int i = 0; i < 27; i++){ printf("%.2f, ", product[i]); } printf("]\n");
    irreps_free(node_sh_irreps);
    irreps_free(neighbor_feature_irreps);

    float weights[] = { 1, 2, 3, 4, 5, 6, 7, 8, 9 };
    //                [ 1 x 1 weight] [1 x 1 weight] [2 x 2 weight] [1 x 1 weight] [1 x 1 weight] [ 1 x 1 weight]
    float output[27] = { 0 };
    Irreps* output_irreps = irreps_create("1x0o + 1x1o + 2x1e + 1x2e + 1x2o + 1x3e");
    linear(product_irreps,
           product,
           weights,
           output_irreps,
           output);

    printf("output ["); for (int i = 0; i < 27; i++) { printf("%.2f, ", output[i]); } printf("]\n");
    irreps_free(product_irreps);
    irreps_free(output_irreps);

    return 0;
}

$ make example && ./example
sh [1.00, 0.46, 0.93, 1.39, 0.83, 0.55, -0.16, 1.66, 1.11, ]
product [13.36, -1.96, 3.93, -1.96, 7.00, 8.00, 9.00, 2.63, 9.50, 16.36, -2.71, 0.00, 4.69, 2.71, -1.36, 9.82, 7.20, -0.38, 13.75, 6.55, 10.76, 13.42, 2.58, -9.40, 5.91, 11.50, 2.93, ]
output [13.36, -3.93, 7.86, -3.93, 24.13, 50.54, 76.95, 30.94, 62.91, 94.88, -18.97, 0.00, 32.86, 18.97, -9.49, 78.56, 57.61, -3.02, 109.98, 52.37, 96.83, 120.75, 23.18, -84.62, 53.18, 103.50, 26.41, ]

Writes the same values to buffer output as the following Python code:

import jax.numpy as jnp
import e3nn_jax as e3nn


node_position = jnp.asarray([1, 2, 3])
node_position_sh = e3nn.spherical_harmonics("1x0e + 1x1o + 1x2e", node_position, normalize=True, normalization="component")
print("sp ", node_position_sh.array)

neighbor_feature = e3nn.IrrepsArray("1x1e", jnp.asarray([7,8,9]))
tp = e3nn.tensor_product(node_position_sh, neighbor_feature)
print("product", tp.array)
linear = e3nn.flax.Linear("1x0o + 1x1o + 2x1e + 1x2e + 1x2o + 1x3e",
                          "1x0o + 1x1o + 2x1e + 1x2e + 1x2o + 1x3e")
weights = {'params': {'w[0,0] 1x0o,1x0o': jnp.asarray([[1]]),
                      'w[1,1] 1x1o,1x1o': jnp.asarray([[2]]),
                      'w[2,2] 2x1e,2x1e': jnp.asarray([[3 , 4], [ 5,  6]]),
                      'w[3,3] 1x2e,1x2e': jnp.asarray([[7]]),
                      'w[4,4] 1x2o,1x2o': jnp.asarray([[8]]),
                      'w[5,5] 1x3e,1x3e': jnp.asarray([[9]])}}
message = linear.apply(weights, tp)
print("output", message.array)

Tetris

See tetris.c which implements a full E(3) equivariant neural network for the classification of tetrominoes. The model can be trained with python train_tetris.py, which saves the model weights to a binary format in tetris.bin. The model can be used for inference by supplying it 4 xyz coordinates on the command line. python run_tetris.py should produce the same outputs using the JAX implementation.

$ make tetris

# usage
$ ./tetris
usage: ./tetris x1 y1 z1 x2 y2 z2 x3 y3 z3 x4 y4 z4

# zigzag
$ ./tetris 0 0 0 1 0 0 1 1 0 2 1 0
logits:
chiral 1    -0.00000
chiral 2    0.00000
square      4.51680
line        1.20807
corner      5.59851
L           4.09760
T           5.82929
zigzag      6.47695

# line 
$ ./tetris 0 0 0 0 0 1 0 0 2 0 0 3
logits:
chiral 1    -0.00000
chiral 2    0.00000
square      0.72002
line        8.12406
corner      -1.34077
L           6.86459
T           4.45846
zigzag      1.23425

# rotated line
$ ./tetris 0 0 0 1 0 0 2 0 0 3 0 0  
logits:
chiral 1    -0.00000
chiral 2    0.00000
square      0.72002
line        8.12406
corner      -1.34077
L           6.86459
T           4.45846
zigzag      1.23425

# line with python 
python run_tetris.py 0 0 0 0 0 1 0 0 2 0 0 3
chiral 1    -0.00000
chiral 2    0.00000
square      0.72002
line        8.12406
corner      -1.34077
L           6.86459
T           4.45846
zigzag      1.23425

Usage

See example above and in message_example.c. Run with

make message_example
./message_example

Currently the output irrep must be defined manually. This could be computed on the fly with minimal computational cost, however I am not sure what makes for the best API here. Additionally, only component normalization is currently implemented, and it will not function properly if the output irreps do not match the full simplified output irreps (i.e. no filtering); see Todo.

Benchmarking

python -m ./venv
source venv/bin/activate
pip install -r extra/requirements.txt

make benchmark

e3nn.c contains several tensor product implementations, each with improvements over the previous for faster runtime.

v1

tensor_product_v1 Is a naive implementation that performs the entire tensor product for all Clebsch-Gordan coefficients:

$$(u \otimes v)^{(l)}_m = \sum_{m_1 = -l_1}^{l_1}\sum_{m_2 = -l_2}^{l_2} C^{(l, m)}_{(l_1, m_1)(l_2, m_2)} u^{(l_1)}_{m_1}v^{(l_2)}_{m_2}$$

To minize overhead in the computation of the Clebsch-Gordan coefficients, they are pre-computed up to L_MAX and cached the first time the tensor product is called, creating a one-time startup cost.

v2

The tensor_product_v2 implementation leverages the fact that, even after conversion to the real basis, the Clebsch-Gordan coeffecients are generally sparse, with many entries equal to 0. To take advantage of this, we precompute a data structure that stores only the non-zero entries of $C$ at each $l_1$, $l_2$, $l$ and their corresponding index at $m_1$, $m_2$, $m$. This significantly improves performance by elminating needless operations of iterating through 0 valued coefficients. Just-in-time (JIT) compilers built into JAX and PyTorch are likely able to perform this optimization as well.

v3

tensor_product_v3 forgoes the computation of Clebsch-Gordan coefficients all together, and instead generates C code to compute the partial tensor product at every $l_1$, $l_2$, $l$ combination up to L_MAX. This elimates the need to iterate over any coefficients, allowing each value in the output to be written in a single step. As it as generated at compile time, the C compliler can also make optimizations to ensure the operations are fast. See tp_codegen.py, which generates tp.c, containing all of the tensor product paths.

Todo:

• Benchmark against e3nn and e3nn-jax
• Sparse Clebsch-Gordan implementation
• Implement Spherical Harmonics
• Implement Linear/Self-interaction operation
• Implement filter_ir_out and irrep_normalization="norm" for tensor product
• Full Nequip, Allegro, or ChargE3Net implementation
• Implement integral, norm, and no normalization for spherical harmonics
• ...

See also

• e3nn PyTorch
• e3nn-jax
• The e3nn paper: https://arxiv.org/abs/2207.09453
• Numerical Recipes in C, 2nd Edition (Press et al.) - helpful formulae and reference implementations for Legendre polynomials, Bessel functions
• karpathy/llama.c - inspo for work