Generating GPU code in TVM (and check out the generated code)

Questions
#1

Hi,

I’m interested in examining some of generated code by TVM for GPU.

Especially, I’d like to generate the convolution code in Fig 15 (for the ResNet-18 model) in the TVM paper (https://www.usenix.org/system/files/osdi18-chen.pdf)

I understand that I need to do some automated search in Fig 12 in the paper.

How can I run the automated search that is similar to Fig 12 in the paper? Any of the two models (ML-based model and Blackbox Genetic Algorithm) would work for me.

If anyone could share the code for that, it would be great! (or anything similar, for example, auto-search code for GEMM is also very helpful.

Thanks!

Recently, I found how i do Genetic algorithm based or ML based automated search(‘autotvm’ module), but i still can’t find method about getting generated cuda codes.

If someone answer the question about getting the cuda codes, it will be very helpful!

#2

You can get the source code generated by tvm in the tvm.Module returned by relay.build.

  • For gpu: source_code = lib.imported_modules[0].get_source()
  • For cpu: source_code = lib.get_source()
    (Where lib is the name of the module returned by relay.build)
1 Like
#3

Hi

I have two more questions

  1. How to specify a customized CUDA codegen when building/compiling the model?
  2. How to output/print the code generated by a customized CUDA codegen?