com.microsoft.EmbedLayerNormalization / build /webgpu /embed-mask-index.wgsl.jinja
Xenova's picture
Xenova HF Staff
sync 6fdf6301e2bb
e774a34 verified
Raw History Blame
1.22 kB
{% macro flat_index_2d(workgroupSize, name="i", bound="params.count", guardInline=false) %}
{% set wgTerm = workgroupSize ~ "u" if workgroupSize is number else workgroupSize %}
// 2D-folded flat index: gid.y carries the high bits past the dispatch's
// per-axis workgroup fold width.
let {{ name }} = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ wgTerm }};
if ({{ name }} >= {{ bound }}) {
return;
}{% endmacro %}
{{ 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>) {
{{ flat_index_2d(maskWorkgroupSize, "batch", "params.batch") }}
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;
}