// SPDX-License-Identifier: AGPL-3.0-or-later
// taxonomy_fc.wgsl — GPU taxonomy Naive Bayes scoring (f64)
//
// wetSpring absorption: metagenomic taxonomy classification.
//
// Computes log-posterior = log_prior + sum(log_prob[feature]) for each taxon.
// One thread per (query, taxon) pair. GEMM-like but with log-space accumulation.
//
// CPU reference: wetspring_barracuda::bio::taxonomy::NaiveBayesClassifier::classify
//
// Bindings:
// 0: config uniform { n_queries, n_taxa, n_features, _pad }
// 1: log_probs [n_taxa × n_features] f64 — log emission probabilities
// 2: log_priors [n_taxa] f64 — log prior probabilities
// 3: features [n_queries × n_features] u32 — binary feature presence
// 4: scores [n_queries × n_taxa] f64 — output log-posteriors
struct TaxConfig {
n_queries: u32,
n_taxa: u32,
n_features: u32,
_pad: u32,
}
@group(0) @binding(0) var<uniform> config: TaxConfig;
@group(0) @binding(1) var<storage, read> log_probs: array<f64>;
@group(0) @binding(2) var<storage, read> log_priors: array<f64>;
@group(0) @binding(3) var<storage, read> features: array<u32>;
@group(0) @binding(4) var<storage, read_write> scores: array<f64>;
@compute @workgroup_size(16, 16)
fn taxonomy_fc(@builtin(global_invocation_id) gid: vec3<u32>) {
let query = gid.x;
let taxon = gid.y;
if query >= config.n_queries || taxon >= config.n_taxa {
return;
}
var score = log_priors[taxon];
let feat_base = query * config.n_features;
let prob_base = taxon * config.n_features;
for (var f: u32 = 0u; f < config.n_features; f = f + 1u) {
let feat_present = features[feat_base + f];
if feat_present != 0u {
score = score + log_probs[prob_base + f];
}
}
scores[query * config.n_taxa + taxon] = score;
}