struct ScalarBatch {
  size: vec4<u32>,
  data: array<f32>,
}

struct Scalar {
  size: vec4<u32>,
  data: array<f32>,
}

struct MatrixBatch {
  size: vec4<u32>,
  data: array<vec4<f32>>,
}

@group(0) @binding(0)
var<storage, read> norm: ScalarBatch;
@group(0) @binding(1)
var<storage, read> threshold: Scalar;
@group(0) @binding(2)
var<storage, read_write> m: MatrixBatch;

@compute @workgroup_size(1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
  m.size = m.size;

  m.data[(global_id.z * m.size.y + global_id.y) * m.size.x / 4 + global_id.x] *= min(1f, threshold.data[0] / norm.data[global_id.z]);
}