Neuronal: modello su ZeroGPU

src/brain.c

@@ -1540,12 +1540,13 @@ int brain_layer_top(Brain *b, int layer, int k, int *idx, float *val, float *mea
 
 // ---- lato remoto (il processo sulla GPU) ----
 
-typedef struct { Brain *b; const BrainLink *l; SBuf s; } ServeCtx;
+typedef struct { Brain *b; const BrainLink *l; SBuf s; double t_digest, t_send; } ServeCtx;
 
 static bool serve_token(void *ud, const char *piece, bool thought) {
     ServeCtx *c = ud;
     Brain *b = c->b;
     SBuf *s = &c->s;
+    double t0 = mono_s();
     s->n = 0;
     uint8_t th = thought;
     uint16_t pl = (uint16_t)strlen(piece);
@@ -1569,7 +1570,11 @@ static bool serve_token(void *ud, const char *piece, bool thought) {
         PUTN(s, b->areas[i].act, sizeof(float) * (size_t)an);
     }
     pthread_mutex_unlock(&b->mu);
-    if (!wire_send(c->l, 'T', s->p, s->n)) return false;
+    double t1 = mono_s();
+    bool sent = wire_send(c->l, 'T', s->p, s->n);
+    c->t_digest += t1 - t0;
+    c->t_send += mono_s() - t1;
+    if (!sent) return false;
     size_t n = 0;
     uint8_t *x = c->l->recv(c->l->ud, &n, 0);           // è arrivato l'ordine di fermarsi?
     bool stop = x && n >= 1 && x[0] == 'X';
@@ -1636,7 +1641,11 @@ void brain_serve(Brain *b, const BrainLink *l, const char *mmproj_path) {
             if (acap > (1u << 22)) acap = 1u << 22;
             char *answer = malloc(acap);
             int32_t res[2];
+            double t0 = mono_s();
+            c.t_digest = c.t_send = 0;
             res[0] = local_chat(b, msgs, nm, &o, (fl & OPF_QUIET) ? NULL : serve_token, &c, answer, acap);
+            fprintf(stderr, "generazione: %d token in %.2f s (riassunto dei neuroni %.2f s, invio %.2f s)" "
+", res[0], mono_s() - t0, c.t_digest, c.t_send);
             res[1] = (int32_t)b->last_stop;
             uint32_t al = res[0] >= 0 ? (uint32_t)strlen(answer) + 1 : 0;
             PUT(&out, res); PUT(&out, al);