// SPDX-License-Identifier: AGPL-3.0-or-later
// Standard Deviation Dim (f64) - Std along a specific dimension at double precision
//
// Similar to PyTorch's torch.std(dim=N)
// Std(X) = sqrt(Var(X)) where Var(X) = E[(X - μ)²]
// Two passes per output element: mean, then variance, then sqrt
struct Params {
dim_size: u32,
outer_size: u32,
inner_size: u32,
}
@group(0) @binding(0) var<storage, read> input: array<f64>;
@group(0) @binding(1) var<storage, read_write> output: array<f64>;
@group(0) @binding(2) 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.outer_size * params.inner_size) {
return;
}
let outer = idx / params.inner_size;
let inner = idx % params.inner_size;
// First pass: compute mean along the dimension
var sum_value = f64(0.0);
for (var i = 0u; i < params.dim_size; i++) {
let input_idx = outer * params.dim_size * params.inner_size + i * params.inner_size + inner;
let value = input[input_idx];
sum_value = sum_value + value;
}
let mean = sum_value / f64(params.dim_size);
// Second pass: compute variance along the dimension
var variance = f64(0.0);
for (var i = 0u; i < params.dim_size; i++) {
let input_idx = outer * params.dim_size * params.inner_size + i * params.inner_size + inner;
let value = input[input_idx];
let diff = value - mean;
variance = variance + diff * diff;
}
variance = variance / f64(params.dim_size);
output[idx] = sqrt(variance);
}