// SPDX-License-Identifier: AGPL-3.0-or-later
// Cross Attention Softmax: Apply softmax to cross-attention scores (f64 canonical)
// Pass 2 of cross-attention
//
// Scores: [batch, heads, decoder_seq, encoder_seq]
// Weights: [batch, heads, decoder_seq, encoder_seq] (after softmax)
//
// For each decoder position i, softmax over all encoder positions j
struct Params {
batch_size: u32,
num_heads: u32,
decoder_seq: u32,
encoder_seq: u32,
head_dim: u32,
}
@group(0) @binding(0) var<storage, read> scores: array<f64>;
@group(0) @binding(1) var<storage, read_write> weights: 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;
let total = params.batch_size * params.num_heads * params.decoder_seq;
if (idx >= total) {
return;
}
// Decompose: [batch, head, decoder_pos]
let b = idx / (params.num_heads * params.decoder_seq);
let h = (idx % (params.num_heads * params.decoder_seq)) / params.decoder_seq;
let i = idx % params.decoder_seq; // decoder position
let base_idx = b * params.num_heads * params.decoder_seq * params.encoder_seq +
h * params.decoder_seq * params.encoder_seq +
i * params.encoder_seq;
// Find max score over all encoder positions
var max_score = f64(-1e10);
for (var j = 0u; j < params.encoder_seq; j++) {
max_score = max(max_score, scores[base_idx + j]);
}
// Compute exp and sum
var sum = f64(0.0);
for (var j = 0u; j < params.encoder_seq; j++) {
let exp_val = exp_f64(scores[base_idx + j] - max_score);
weights[base_idx + j] = exp_val;
sum += exp_val;
}
// Normalize
for (var j = 0u; j < params.encoder_seq; j++) {
weights[base_idx + j] = weights[base_idx + j] / sum;
}
}