TPUMLIR main objectives and development plans

This document lists general directions of TPU-MLIR, include principal goals, plans, etc.

Principal goals#

  • keep regular sync with mlir from project llvm-project.
  • continue maintaining high-quality, well-tested and documented modules.
  • support AI machine learning frameworks: ONNX, TFLite, Caffe.
  • support various neuron network models, such as resnet, yolo, ssd, bert, etc.
  • support various chips, especially SOPHGO TPU, such as BM1684X, BM1684, CV18XX, Athena2, etc.
  • follow newest MLIR compiler technology, and put it on this project efficently.
  • keep high performance and hign accuraccy, especially for INT8 quantization models.
  • support various tools to improve development conveniently and efficiently.

Plans#

development plans【2022-2023】#

Notes:
Chip Product
BM1684X SC7, SE7
BM1684 SC5, SG6, SE6, SE5, SM5
CV18XX CV1838, CV1835, CV1826, CV1825, CV1823, CV1821, CV1820, CV1813H, CV1812H, CV1811H, CV1810H, CV1812C, CV1811C
Athena2 on going