audio: tensorflow: tune: export model blobs for topology and sof-ctl

Update sof_tflm_train.py to package the trained tflite model behind
the SOF IPC4 ABI header (struct sof_abi_hdr) and export it to:

  - tools/topology/topology2/include/components/tflm/*.conf for
    embedding into ALSA topology v2 files.
  - tools/ctl/ipc4/tflm/*.txt for runtime application with sof-ctl.

Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>

Author: singalsu

src/audio/tensorflow/tune/sof_tflm_train.py

@@ -30,7 +30,9 @@
 from __future__ import annotations
 
 import argparse
+import datetime
 import os
+import struct
 import sys
 from pathlib import Path
 
@@ -131,6 +133,24 @@ def convert_to_tflite_int8(
     return converter.convert()
 
 
+def set_tflite_model_description(tflite_bytes: bytes, description: str) -> bytes:
+    """Embed comma-separated labels into TFLite FlatBuffer Model.description."""
+    try:
+        from tensorflow.lite.python import schema_py_generated as schema_fb
+        import flatbuffers
+
+        model_obj = schema_fb.ModelT.InitFromObj(
+            schema_fb.Model.GetRootAsModel(tflite_bytes, 0)
+        )
+        model_obj.description = description
+        builder = flatbuffers.Builder(len(tflite_bytes) + len(description) + 256)
+        builder.Finish(model_obj.Pack(builder), "TFL3")
+        return bytes(builder.Output())
+    except Exception as exc:
+        print(f"    [WARN] could not embed description in tflite FlatBuffer: {exc}")
+        return tflite_bytes
+
+
 def emit_c_array(
     tflite_bytes: bytes,
     out_cc: Path,
@@ -183,6 +203,69 @@ def emit_labels_header(labels: list[str], out_h: Path) -> None:
         f.write(f"#endif  // {guard}\n")
 
 
+# SOF IPC4 ABI definitions for binary control blobs
+SOF_IPC4_ABI_MAGIC = 0x34464F53  # 'SOF4' in little-endian
+SOF_ABI_VERSION = (3 << 24) | (29 << 12) | 1  # 3.29.1 (0x0301d001)
+SOF_CTRL_CMD_BINARY = 3
+
+
+def build_ipc4_abi_blob(tflite_bytes: bytes, param_id: int = 0) -> bytes:
+    """Pack tflite model bytes behind a standard 32-byte struct sof_abi_hdr."""
+    size = len(tflite_bytes)
+    abi_header = struct.pack("<IIIIIIII", SOF_IPC4_ABI_MAGIC, param_id, size, SOF_ABI_VERSION, 0, 0, 0, 0)
+    pad = (4 - (size % 4)) % 4
+    return abi_header + tflite_bytes + b"\x00" * pad
+
+
+def emit_topology2_conf(
+    tflite_bytes: bytes,
+    out_conf: Path,
+    model_name: str,
+    labels: list[str],
+) -> None:
+    """Emit Topology2 configuration blob in .conf format for inclusion to topology."""
+    blob8 = build_ipc4_abi_blob(tflite_bytes)
+    words = struct.unpack(f"<{len(blob8) // 4}I", blob8)
+    out_conf.parent.mkdir(parents=True, exist_ok=True)
+    today = datetime.date.today().strftime("%d-%b-%Y")
+    labels_str = ", ".join(labels)
+    data_name = f"tflm_config_{model_name}"
+    with open(out_conf, "w") as f:
+        f.write(f"# Exported TFLM Model Control Words {today}\n")
+        f.write(f"# Model: {model_name} (classes: {labels_str})\n")
+        f.write(f"# Auto-generated by sof_tflm_train.py — do not edit.\n")
+        f.write(f'Object.Base.data."{data_name}" {{\n')
+        f.write('\twords "\n')
+        lines = []
+        for i in range(0, len(words), 8):
+            chunk = words[i : i + 8]
+            lines.append(",".join(f"0x{w:08x}" for w in chunk))
+        for idx, line in enumerate(lines):
+            if idx == len(lines) - 1:
+                f.write(f"\t\t{line}\"\n")
+            else:
+                f.write(f"\t\t{line},\n")
+        f.write("}\n")
+
+
+def emit_sofctl_ipc4_txt(
+    tflite_bytes: bytes,
+    out_txt: Path,
+) -> None:
+    """Emit sof-ctl IPC4 text blob in comma-separated uint32 CSV format."""
+    size = len(tflite_bytes)
+    tlv_cmd = SOF_CTRL_CMD_BINARY
+    tlv_size = 32 + size
+    pad = (4 - (size % 4)) % 4
+    payload_padded = tflite_bytes + b"\x00" * pad
+    words = [tlv_cmd, tlv_size, SOF_IPC4_ABI_MAGIC, 0, size, SOF_ABI_VERSION, 0, 0, 0, 0]
+    payload_words = list(struct.unpack(f"<{len(payload_padded) // 4}I", payload_padded))
+    all_words = words + payload_words
+    out_txt.parent.mkdir(parents=True, exist_ok=True)
+    with open(out_txt, "w") as f:
+        f.write(",".join(str(w) for w in all_words) + "\n")
+
+
 def parse_args() -> argparse.Namespace:
     ap = argparse.ArgumentParser(description=__doc__)
     ap.add_argument("--feat-root", required=True)
@@ -230,11 +313,12 @@ def parse_args() -> argparse.Namespace:
 
 def main() -> int:
     args = parse_args()
+    reserved = {"silence", "unknown"}
+    keywords = [lbl for lbl in args.labels if lbl not in reserved]
+
     if not args.name:
-        # Derive a sensible default: first keyword label past silence/unknown.
-        reserved = {"silence", "unknown"}
-        keywords = [lbl for lbl in args.labels if lbl not in reserved]
-        args.name = keywords[0] if keywords else "wov_model"
+        # Default name contains all trained keywords joined with underscore
+        args.name = "_".join(keywords) if keywords else "wov_model"
         print(f">>> --name not set, using {args.name!r}")
     tf.keras.utils.set_random_seed(args.seed)
 
@@ -284,6 +368,7 @@ def main() -> int:
 
     print(">>> Converting to int8 tflite")
     tflite_bytes = convert_to_tflite_int8(model, X_tr, y_tr, args.rep_samples)
+    tflite_bytes = set_tflite_model_description(tflite_bytes, ",".join(args.labels))
 
     out_dir = Path(args.out_dir)
     out_dir.mkdir(parents=True, exist_ok=True)
@@ -309,6 +394,36 @@ def main() -> int:
     labels_path.write_text("\n".join(args.labels) + "\n")
     print(f"    wrote {labels_path} (archive copy of the label list)")
 
+    # Export Topology2 config blob (.conf) and sof-ctl IPC4 text blob (.txt)
+    out_conf = out_dir / f"{args.name}.conf"
+    emit_topology2_conf(tflite_bytes, out_conf, args.name, args.labels)
+    print(f"    wrote {out_conf}")
+
+    out_txt = out_dir / f"{args.name}.txt"
+    emit_sofctl_ipc4_txt(tflite_bytes, out_txt)
+    print(f"    wrote {out_txt}")
+
+    # Also export directly to SOF tools directories if present in workspace/repo
+    script_dir = Path(__file__).resolve().parent
+    repo_root = None
+    for parent in script_dir.parents:
+        if (parent / "tools/topology/topology2").is_dir():
+            repo_root = parent
+            break
+    if repo_root is None and "SOF_WORKSPACE" in os.environ:
+        candidate = Path(os.environ["SOF_WORKSPACE"]) / "sof"
+        if (candidate / "tools/topology/topology2").is_dir():
+            repo_root = candidate
+
+    if repo_root is not None:
+        tplg_tflm_dir = repo_root / "tools/topology/topology2/include/components/tflm"
+        emit_topology2_conf(tflite_bytes, tplg_tflm_dir / f"{args.name}.conf", args.name, args.labels)
+        print(f"    exported {tplg_tflm_dir / f'{args.name}.conf'}")
+
+        ctl_tflm_dir = repo_root / "tools/ctl/ipc4/tflm"
+        emit_sofctl_ipc4_txt(tflite_bytes, ctl_tflm_dir / f"{args.name}.txt")
+        print(f"    exported {ctl_tflm_dir / f'{args.name}.txt'}")
+
     return 0
 
 

tools/ctl/ipc4/tflm/banana_strawberry_orange.txt

@@ -19,6 +19,14 @@ Class.Base."data" {
 		type	"string"
 	}
 
+	DefineAttribute."shorts" {
+		type	"string"
+	}
+
+	DefineAttribute."words" {
+		type	"string"
+	}
+
 	attributes {
 		!constructor [
 			"name"