[BACKEND][CODEGEN] C codegen with tests

[mutinifni]

This PR introduces a gcc / g++ compatible C code generator for TVM, tested on x86 host.

[tqchen]

@Huyuwei can you help review this?

[Huyuwei]

@mutinifni Can you briefly describe the current compilation flow? Most importantly, is it using DSOModule for the runtime?

[mutinifni]

@Huyuwei It uses CSourceModule for saving the C source code - I have verified this by checking the type_key. Loading is done via the DSOModule.

I was a bit uncertain about this module design and the export_library functionality as well, so please feel let me know if you think a different flow would be better instead.

[Huyuwei]

replace "__tvm_module_ctx" with tvm::runtime::symbol::tvm_module_ctx?

[mutinifni]

This would make the generated code necessarily dependent on C++, but we want to keep the option of pure C as well. If there's a reason to use the latter symbol, I can change it.

[tqchen]

@Huyuwei please take another look

[srkreddy1238]

This PR cause the below issue with llvm target.

-tgt="cuda"
+tgt="llvm"

python3 tutorials/get_started.py

!115 = !{!116, !116, i64 0}
!116 = !{!"float32", !117, i64 0}
!117 = !{!"0x119a190", !4, i64 0}
Traceback (most recent call last):
File "get_started.py", line 196, in <module>
cc.create_shared(temp.relpath("myadd.so"), [temp.relpath("myadd.o")])
File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/contrib/cc.py", line 34, in create_shared
_linux_shared(output, objects, options, cc)
File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/contrib/cc.py", line 53, in _linux_shared
cmd += ["-I" + path for path in find_include_path()]
File "/home/srk/.local/lib/python3.6/site-packages/tvm-0.5.dev0-py3.6-linux-x86_64.egg/tvm/_ffi/libinfo.py", line 156, in find_include_path
raise RuntimeError(message)
RuntimeError: Cannot find the files.
List of candidates:
/home/srk/.local/lib/python3.6/include
/home/srk/.local/lib/python3.6/site-packages/include
/home/srk/.local/lib/python3.6/site-packages/3rdparty/include
/home/srk/.local/lib/python3.6/dlpack/include
/home/srk/.local/lib/python3.6/site-packages/dlpack/include
/home/srk/.local/lib/python3.6/site-packages/3rdparty/dlpack/include

[tqchen]

Thanks @srkreddy1238 for reporting this, @mutinifni can you send in a quick fix? We should not directly raise error when the headers are not found in general create_shared, instead, we should only call find_include_path and add it to the options in created_shared when it is a c module

[dongxiao92]

Hello, is there any documents or examples about how to use this c backend?

[mutinifni]

Hi @dongxiao92, you can look at tests/python/unittest/test_codegen_c_host.py for a vector addition example which runs on the host. You can also print out the generated C code from the function export_library in python/tvm/module.py.

[wujiaqi6630]

Hello, I want to generate C code too:

func = tvm.build(s, [A, B, C], target=device, name='myGemm')
temp = util.tempdir()
path_dso = temp.relpath("temp.so")
path = temp.relpath('lib.tar')
func.export_library(path_dso)
m = tvm.module.load(path_dso)
print(m.get_source())

But failed:
TVMError: Module[dso] does not support GetSource
Am I using the wrong method? @mutinifni

[mutinifni]

Hi @wujiaqi6630, if you want to see the C source code that is generated, you can use func.get_source() because func is the C source module where as m is the DSO module. Only the C source module supports GetSource.

[wujiaqi6630]

Thank you !

[Jie-KUN]

Hi@Mutinifni, I use the func.get_source(), but the code printed on the screen seems not to run on the X86 platform. How can I execute the C code. Looking forward to your reply

[NiteshNarayana]

Hi

Any update on how to execute the generated C source code?

I observe that when you generate the c source code for a module that is generated through a user-defined relay function you get a __tvm_main__() function generated as well. However this is not the case when I generate the source code for a network, eg mnist, where I see a different set of functions (eg. tvmgen_default_fused_nn_conv2d_add_nn_relu , tvmgen_default_fused_nn_max_pool2d_1) which I suppose are one for each layer. How are we to explicitly compile this and run it?

Thanks in advance!

[shreyashvi]

Hi

Any update on how to execute the generated C source code?

I observe that when you generate the c source code for a module that is generated through a user-defined relay function you get a __tvm_main__() function generated as well. However this is not the case when I generate the source code for a network, eg mnist, where I see a different set of functions (eg. tvmgen_default_fused_nn_conv2d_add_nn_relu , tvmgen_default_fused_nn_max_pool2d_1) which I suppose are one for each layer. How are we to explicitly compile this and run it?

Thanks in advance!

Hey, did you find out a way how to compile that generated c code?

Thanks