add throughput mode to validation scripts

make raptor also emit input sizes

validation/raptor_validation/onnx_utils.py

@@ -156,6 +156,42 @@ def gen_random_inputs(
     return arrays_in_order, arrays_by_name
 
 
+def generate_input_batch(onnx_inputs, first_inputs, batch_size, seed):
+    if batch_size < 1:
+        raise ValueError("batch size must be at least 1")
+    if not onnx_inputs:
+        return [first_inputs] * batch_size
+
+    batch = [first_inputs]
+    for index in range(1, batch_size):
+        sample, _ = gen_random_inputs(onnx_inputs, seed=seed + index)
+        if all(np.array_equal(left, right) for left, right in zip(sample, batch[-1])):
+            sample[0] = sample[0].copy()
+            if sample[0].size == 0:
+                raise ValueError("throughput validation cannot distinguish empty input tensors")
+            if np.issubdtype(sample[0].dtype, np.bool_):
+                sample[0].flat[0] = not sample[0].flat[0]
+            elif np.issubdtype(sample[0].dtype, np.integer):
+                info = np.iinfo(sample[0].dtype)
+                value = sample[0].flat[0]
+                sample[0].flat[0] = value + 1 if value < info.max else value - 1
+            else:
+                sample[0].flat[0] += 1
+        batch.append(sample)
+    return batch
+
+
+def write_input_batch_csv(path, input_batch):
+    path = pathlib.Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with path.open("w", newline="", encoding="utf-8") as output:
+        writer = csv.writer(output)
+        for sample in input_batch:
+            writer.writerow(
+                np.concatenate([array.reshape(-1) for array in sample]) if sample else ()
+            )
+
+
 def save_inputs_to_files(onnx_path, arrays_in_order, out_dir):
     """
     Save arrays to CSV files. Returns (flags, files) where flags is a list
@@ -201,3 +237,22 @@ def write_inputs_to_memory_bin(memory_bin_path, config_json_path, arrays_in_orde
             native = arr.astype(arr.dtype.newbyteorder("="), copy=False)
             f.seek(addr)
             f.write(native.tobytes(order="C"))
+
+
+def write_inputs_binary(path, arrays_in_order):
+    """Write one simulator input in graph-input order."""
+    with open(path, "wb") as f:
+        for arr in arrays_in_order:
+            native = arr.astype(arr.dtype.newbyteorder("="), copy=False)
+            f.write(native.tobytes(order="C"))
+
+
+def write_input_batch_binaries(input_batch, output_dir, transform=None):
+    output_dir = pathlib.Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    paths = []
+    for index, sample in enumerate(input_batch):
+        path = output_dir / f"input_{index}.bin"
+        write_inputs_binary(path, [transform(sample[0])] if transform is not None else sample)
+        paths.append(path)
+    return paths