TIDL is TI's software ecosystem for Deep learning algorithm (CNN) acceleration.

TIDL allows users to run inference for pre-trained CNN/DNN models on TI Devices with C6x or C7x DSP. Examples showing usage of TIDL are provided as part of Processor SDK RTOS Automotive.

Fundamental blocks of TI Deep Learning Library

User can deploy the CNN application using one of below options

The openVX based API in A72 to execute complete CNN model on C7x-MMA
Tensorflow Lite runtime based Heterogeneous Execution on A72+C7x-MMA
- Offloads subgraphs to C7x/MMA for accelerated execution via TFLite Delgate API
- Execution on ARM core for layers that are not supported in C7x-MMA
Onnx runtime based Heterogeneous Execution on A72+C7x-MMA
- Offloads subgraphs to C7x/MMA for accelerated execution via TIDL Execution Provider in onnx runtime
- Execution on ARM core for layers that are not supported in C7x-MMA
TVM/Neo-AI-DLR based Heterogeneous Execution on A72+C7x-MMA
- Offloads subgraphs to C7x/MMA for accelerated execution
- Execution on ARM core for layers that are not supported in C7x-MMA

Refer Tensorflow Lite runtime + TIDL Heterogeneous Execution for more details on TFLite runtime based programming model.

Refer Onnx runtime + TIDL Heterogeneous Execution for more details on Onnx runtime based programming model.

Refer TVM/Neo-AI-DLR + TIDL Heterogeneous Execution for more details on TVM/Neo-AI-DLR programming model.