This is a fork of TVM for adding BYOC integrations for the 3LA project.
Right now we have a VTA integration in src/relay/backend/contrib/vta_matmul. Note that you have to include the line SET(USE_VTA_MATMUL ON) in build/config.cmake before building TVM to support this (other flags that should be on: USE_LLVM, USE_VTA_FSIM). We have a test of this backend in tests/python/relay/test_external_codegen.py (see test_extern_vta()).
This version also uses a fork of the VTA repo meant to dump logs.
Try vta/python/integration/matmul_tutorial.py to use the dumping facility.
VTA can be set into dumping mode by calling vta.testing.simulator.dump_mode(True).
You can specify the location at which the dump will be deposited using vta.testing.simulator.dump_target(path); the default is ./vta_sim_dump.json.
See the readme at the VTA fork to see a description of the dumping mode and the dumping format.
You can use vta.testing.ila_converter.convert(dump_file, dest_file) to convert a VTA simulator dump into an ILA program fragment.
Open Deep Learning Compiler Stack
Documentation |
Contributors |
Community |
Release Notes
Apache TVM (incubating) is a compiler stack for deep learning systems. It is designed to close the gap between the productivity-focused deep learning frameworks, and the performance- and efficiency-focused hardware backends. TVM works with deep learning frameworks to provide end to end compilation to different backends.
© Contributors Licensed under an Apache-2.0 license.
TVM adopts apache committer model, we aim to create an open source project that is maintained and owned by the community. Checkout the Contributor Guide
We learned a lot from the following projects when building TVM.
- Halide: Part of TVM's TIR and arithmetic simplification module originates from Halide. We also learned and adapted some part of lowering pipeline from Halide.
- Loopy: use of integer set analysis and its loop transformation primitives.
- Theano: the design inspiration of symbolic scan operator for recurrence.