Function process_logits

tools/perplexity/perplexity.cpp:109–142

}

static void process_logits(
    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
    double & nll, double & nll2, float * logit_history, float * prob_history
) {
    std::mutex mutex;
    int counter = 0;
    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
        double local_nll  = 0;
        double local_nll2 = 0;
        while (true) {
            std::unique_lock<std::mutex> lock(mutex);
            int i = counter++;
            if (i >= n_token) {
                nll += local_nll; nll2 += local_nll2;
                break;
            }
            lock.unlock();
            const results_log_softmax results = log_softmax(n_vocab, logits + size_t(i)*n_vocab, tokens[i+1]);
            const double v = -results.log_softmax;
            local_nll += v;
            local_nll2 += v*v;

            logit_history[i] = results.logit;
            prob_history[i]  = results.prob;
        }
    };
    for (auto & w : workers) {
        w = std::thread(compute);
    }
    compute();
    for (auto & w : workers) {
        w.join();
    }
}

static void process_logits(std::ostream& out, int n_vocab, const float * logits, const int * tokens, int n_token,
        std::vector<std::thread> & workers, std::vector<uint16_t> & log_probs, double & nll, double & nll2) {