How do I use the code I generate

#1

I used tvm to translate a resnet18 I built in Python into C language. However, the generated C language code is extremely complicated, and there seems to be no top-level interface. How should I use the generated C code?

#2

complicated as it is indeed, it’s ok if we can build it with no error(warnings could be accepted!) :smiley:

1 Like
#3

大哥，生成出来的是这种，下一步该怎么整 :sob:

36259883de9d27cd08551589b3af3cf
36259883de9d27cd08551589b3af3cf1718×918 219 KB

#4

我编译了一下，但是很多报错 :sob:

7f8efa26bba0cf70710b7da796c4d36
7f8efa26bba0cf70710b7da796c4d361718×918 268 KB

#5

try surrounding your relay.build with

with tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True}):
    relay.build(...)
#6

TVM codegen only works for operators but not the entire graph. So it’s impossible to run the e2e model via c code. (I am not sure if AOT helps, not familiar with it)

TVM uses a runtime to execute the model if you compile it via relay or relax

#7

Thank you sir, I successfully compiled it！ :smiley_cat:

#8

Got you bro. But what I want to do is, can I convert the model in Python into a model in C and run it correctly under C?This is a project my teacher asked me to do, and I know very little about compilers,so I just try to do in this way. Does this “for operators” mean to use lamba expressions to translate operators not the whole model?

#9

生成的op的C代码，需要用relay(or relax)编译，并加载到虚拟机里执行，如果你用的是异构设备，还涉及到Host端代码。总之，即使你想跑一下很简单的z=x + y，你无法脱离这个编译-推理架构，我理解你应该是想生成纯C代码，使用GCC编译，生成可执行文件(gcc -o a.out xxx.cc， and ./a.out)

1 Like
#10

能加个vx吗大哥 :sob:有点问题想请教一下您