// SPDX-License-Identifier: AGPL-3.0-or-later
// Scatter - Write values to specific indices (f64 canonical)
//
// Deep Debt Principles:
// - Pure WGSL implementation (universal compute)
// - Zero unsafe code (memory safe)
// - Hardware-agnostic (works on any GPU/CPU via WebGPU)
// - Self-contained logic (no external dependencies)
struct Params {
size: u32,
dim_size: u32,
outer_size: u32,
inner_size: u32,
scatter_size: u32,
}
@group(0) @binding(0) var<storage, read> input: array<f64>;
@group(0) @binding(1) var<storage, read> indices: array<u32>;
@group(0) @binding(2) var<storage, read> values: array<f64>;
@group(0) @binding(3) var<storage, read_write> output: array<f64>;
@group(0) @binding(4) var<uniform> params: Params;
@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let idx = global_id.x;
if (idx >= params.size) {
return;
}
// First, copy input to output
output[idx] = input[idx];
}
// Second pass: scatter values
@compute @workgroup_size(256)
fn scatter(@builtin(global_invocation_id) global_id: vec3<u32>) {
let idx = global_id.x;
if (idx >= params.outer_size * params.scatter_size * params.inner_size) {
return;
}
// Decompose scatter index
let outer = idx / (params.scatter_size * params.inner_size);
let mid = (idx / params.inner_size) % params.scatter_size;
let inner = idx % params.inner_size;
// Get the index to scatter to
let scatter_idx = indices[mid];
// Bounds check
if (scatter_idx >= params.dim_size) {
return;
}
// Calculate output position
let output_idx = outer * params.dim_size * params.inner_size +
scatter_idx * params.inner_size + inner;
// Write the value (atomic not needed in WebGPU for non-overlapping writes)
output[output_idx] = values[idx];
}