Kernels:

---

webgpu-kernels

/

com.microsoft.EmbedLayerNormalization

like 0

Follow

WebGPU Kernels Community 15

{{ env.wgsl.resourceDeclarations }}

// com.microsoft.EmbedLayerNormalization, mask_index pass. With a mask, return
// the first zero position or the sequence length when every position is set.
// Without a mask, initialize the optional output to zero.
{% if hasMask %}
const SEQUENCE: u32 = {{ sequenceLength }}u;
{% endif %}

@compute @workgroup_size({{ maskWorkgroupSize }}, 1, 1)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
  let batch = gid.x;
  if (batch >= params.batch) {
    return;
  }
  var first_zero: i32 = 0;
{% if hasMask %}
  first_zero = i32(SEQUENCE);
  let base = batch * SEQUENCE;
  for (var s: u32 = 0u; s < SEQUENCE; s = s + 1u) {
    if (mask[base + s] == 0) {
      first_zero = i32(s);
      break;
    }
  }
{% endif %}
  mask_index[batch] = first_zero;
}