TensorFlow Lite NIF for Mob apps.

Loads .tflite model bytes, runs inference via TFLite's bundled XNNPACK CPU path or via the NNAPI delegate (which on MediaTek devices like the Moto G Power 5G 2024 routes to the mtk-gpu_shim GPU accelerator and gets us ~150 ms YOLOv8n).

Example

tflite = File.read!("priv/yolov8n_full_integer_quant.tflite")
{:ok, m} = NxTfliteMob.load_module(tflite,
             delegate: "nnapi",
             accelerator: "mtk-gpu_shim",
             allow_fp16: true)

input_int8 = File.read!("priv/input_int8.bin")  # 1x640x640x3 INT8
{:ok, [out_bin]} = NxTfliteMob.call(m, [input_int8])

:ok = NxTfliteMob.release_module(m)

Delegates

• :delegate — "xnnpack" (default, CPU INT8/FP32) or "nnapi" (vendor NN HAL, hits the GPU / NPU when present)
• :accelerator — only meaningful with delegate: "nnapi"; the accelerator name to request. Discover with NxTfliteMob.NIF.list_nnapi_devices/0 (planned). Known values on the Moto BXM-8-256:
  • "mtk-gpu_shim" — PowerVR GPU through MediaTek NNAPI HAL (best result for YOLOv8n)
  • "mtk-neuron_shim" — APU NPU; only partial op coverage for YOLO, falls back partial
  • "nnapi-reference" — NNAPI's CPU emulation (slow)
• :num_threads — XNNPACK CPU thread count (default 6)
• :allow_fp16 — NNAPI may run FP32 ops in FP16 (default true)