bench : fix timings by running a pre-heat

examples/bench/bench.cpp

@@ -70,6 +70,15 @@ int whisper_bench_encoder(const whisper_params & params) {
         return 3;
     }
 
+    // heat up
+    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        return 4;
+    }
+
+    whisper_reset_timings(ctx);
+
+    // actual run
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
         fprintf(stderr, "error: failed to encode model: %d\n", ret);
         return 4;

extra/bench-all.sh

@@ -48,9 +48,6 @@ printf "| CPU | OS | Config | Model | Th | Load | Enc. | Commit |\n"
 printf "| --- | -- | ------ | ----- | -- | ---- | ---- | ------ |\n"
 
 for model in "${models[@]}"; do
-    # run once to heat-up the cache
-    ./bench -m ./models/ggml-$model.bin -t $n_threads 2>/dev/null 1>/dev/null
-
     # actual run
     # store stderr output in a variable in order to parse it later
     output=$(./bench -m ./models/ggml-$model.bin -t $n_threads 2>&1)

whisper.cpp

@@ -3588,6 +3588,9 @@ void whisper_reset_timings(struct whisper_context * ctx) {
         ctx->state->t_sample_us = 0;
         ctx->state->t_encode_us = 0;
         ctx->state->t_decode_us = 0;
+        ctx->state->n_sample = 0;
+        ctx->state->n_encode = 0;
+        ctx->state->n_decode = 0;
     }
 }