primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
Documentation
// SPDX-License-Identifier: AGPL-3.0-or-later
// GATConv - Graph Attention Networks (f64 canonical)

struct Params {
    num_nodes: u32,
    num_edges: u32,
    in_features: u32,
    out_features: u32,
    leaky_slope: f64,
    _pad1: u32,
    _pad2: u32,
    _pad3: u32,
}

@group(0) @binding(0) var<uniform> params: Params;
@group(0) @binding(1) var<storage, read> node_features: array<f64>;
@group(0) @binding(2) var<storage, read> edge_index: array<u32>;
@group(0) @binding(3) var<storage, read> weights: array<f64>;
@group(0) @binding(4) var<storage, read> attention: array<f64>;
@group(0) @binding(5) var<storage, read_write> transformed: array<f64>;
@group(0) @binding(6) var<storage, read_write> output: array<atomic<i32>>;

@compute @workgroup_size(256)
fn transform_features(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let node = global_id.x;
    if (node >= params.num_nodes) {
        return;
    }

    for (var out_f = 0u; out_f < params.out_features; out_f = out_f + 1u) {
        var sum: f64 = 0.0;
        for (var in_f = 0u; in_f < params.in_features; in_f = in_f + 1u) {
            let feat_idx = node * params.in_features + in_f;
            let weight_idx = in_f * params.out_features + out_f;
            sum += node_features[feat_idx] * weights[weight_idx];
        }
        transformed[node * params.out_features + out_f] = sum;
    }
}

@compute @workgroup_size(256)
fn aggregate(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let edge = global_id.x;
    if (edge >= params.num_edges) {
        return;
    }

    let src = edge_index[edge * 2u];
    let dst = edge_index[edge * 2u + 1u];

    var score: f64 = 0.0;
    for (var f = 0u; f < params.out_features; f = f + 1u) {
        let src_val = transformed[src * params.out_features + f];
        let dst_val = transformed[dst * params.out_features + f];
        score += src_val * attention[f];
        score += dst_val * attention[params.out_features + f];
    }

    var alpha: f64;
    if (score > 0.0) {
        alpha = score;
    } else {
        alpha = params.leaky_slope * score;
    }
    let weight = exp_f64(alpha);

    for (var f = 0u; f < params.out_features; f = f + 1u) {
        let val = weight * transformed[src * params.out_features + f];
        atomicAdd(&output[dst * params.out_features + f], bitcast<i32>(f32(val)));
    }
}