In paper, micro-kernels compiled by JIT are computed in PS? GEMM and ALU are in PL?

Questions
#1

I’m a student and new in this.
I read the paper TVM: An Automated End-to-End Optimizing Compiler for Deep Learning, it mentioned that TVM use FPGA for acceleration. My understanding is that the GEMM and ALU are in PL side, but the micro-kernels compiled by JIT for flexibility are in PS, because it is generate by software but not bitstream. Is it right?
If above is correct, I didn’t find ***.bit file in github project TVM/vta/ . Do we need to generate bit stream ourselves with vivado?
Thanks a lot~

#2

Yes and it is called VTA, you can find more information here

https://docs.tvm.ai/vta/index.html

#3

Thank you for your reply;
I read the document but two things still confuse me;
1.It told me new user can’t put image here.
This page demo search “bitstream”: https://docs.tvm.ai/vta/tutorials/frontend/deploy_resnet_on_vta.html#sphx-glr-vta-tutorials-frontend-deploy-resnet-on-vta-py
in this demo, why it program_fpga with none bitstream. The LOAD, COMPUTE, STORE module is in FPGA, how it get the hardware information;
2. this page search “The runtime can readily make use of schedules to” :https://arxiv.org/pdf/1807.04188.pdf

how the mocro-kernels compiled by JIT works? My understanding is that it turn the new operators into a fix of GEMM and ALU for computing. Is it right?

#4

Hey,

Can you elaborate a little bit more on (1).

Regarding (2), you can find more information about VTA runtime and how it is used in the VTA pass

#5

(1) if you set it to None, it will by download the default bitstream for the vta parameterization. You can answer your question by looking at the definition of that function.

#6

(2) microcode generation is handled by the runtime; if you want to see how these are generated you can add with vta.build_config(debug_flag = 0x6): to the simpler tutorial examples before calling tvm.build()

for instance in the simple matrix multiply tutorial, you can use

with vta.build_config(debug_flag = 0x6):
    my_gemm = tvm.build(s, [A, B, C], "ext_dev",
                        env.target_host, name="my_gemm")

and that will get the runtime to print out the microcoded kernels that get generated

1 Like