Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add Initial Cross-Compilation Documentation for RISC-V Support #19239

Merged

Conversation

phoebesv
Copy link
Contributor

@phoebesv phoebesv commented Jan 23, 2024

Update the document for PR: Enable Cross-Compiling Support for ONNX Runtime on Linux for RISC-V 64-bit.
#19238

snnn pushed a commit that referenced this pull request Jan 25, 2024
#19238)

### Description  
This pull request introduces the necessary changes to enable RISC-V
64-bit cross-compiling support for the ONNX Runtime on Linux. The RISC-V
architecture has gained popularity as an open standard instruction set
architecture, and this contribution aims to extend ONNX Runtime's
compatibility to include RISC-V, thereby broadening the reach of ONNX
models to a wider range of devices.

### Motivation and Context
RISC-V is a free and open-source instruction set architecture (ISA)
based on established RISC principles. It is provided under open licenses
without fees. Due to its extensibility and freedom in both software and
hardware, RISC-V is poised for widespread adoption in the future,
especially in applications related to AI, parallel computing, and data
centers.

### Example Build Command
```
./build.sh --parallel --config Debug --rv64 --riscv_toolchain_root=/path/to/toolchain/root --skip_tests
```

### Documentation Updates
Relevant sections of the documentation will be updated to reflect the
newly supported RISC-V 64-bit cross-compilation feature.
#19239

---------

Signed-off-by: Phoebe Chen <[email protected]>
Copy link
Contributor

@yihonglyu yihonglyu left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@phoebesv I followed the instructions but encounter the error message

[ 51%] Building CXX object CMakeFiles/onnxruntime_providers_xnnpack.dir/home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc.o
/home/yilyu/risc-v/riscv/bin/riscv64-unknown-linux-gnu-g++ --sysroot=/home/yilyu/risc-v/riscv/sysroot -DEIGEN_MPL2_ONLY -DEIGEN_USE_THREADS -DNSYNC_ATOMIC_CPP11 -DONNX_ML=1 -DONNX_NAMESPACE=onnx -DONNX_USE_LITE_PROTO=1 -DORT_ENABLE_STREAM -DPLATFORM_POSIX -DUSE_XNNPACK=1 -D_GNU_SOURCE -D__ONNX_NO_DOC_STRINGS -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/utf8_range-src -I/home/yilyu/onnxruntime/include/onnxruntime -I/home/yilyu/onnxruntime/include/onnxruntime/core/session -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/google_nsync-src/public -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo -I/home/yilyu/onnxruntime/onnxruntime -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/abseil_cpp-src -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/safeint-src -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/gsl-src/include -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/onnx-src -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/onnx-build -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/protobuf-src/src -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/googlexnnpack-src/include -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/pthreadpool-src/include -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/flatbuffers-src/include -I/home/yilyu/onnxruntime/build/Linux/RelWithDebInfo/risc-v/RelWithDebInfo/_deps/mp11-src/include -DNDEBUG -Wp,-D_FORTIFY_SOURCE=2 -Wp,-D_GLIBCXX_ASSERTIONS -fstack-protector-strong -O3 -pipe -ggdb3 -fstack-clash-protection -ffunction-sections -fdata-sections -Wno-error=attributes -Wno-restrict -O2 -g -DNDEBUG -fPIC -Wall -Wextra -Wno-deprecated-copy -Wno-nonnull-compare -Werror -MD -MT CMakeFiles/onnxruntime_providers_xnnpack.dir/home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc.o -MF CMakeFiles/onnxruntime_providers_xnnpack.dir/home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc.o.d -o CMakeFiles/onnxruntime_providers_xnnpack.dir/home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc.o -c /home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc
In file included from /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/string:51,
from /home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.h:6,
from /home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc:4:
In static member function 'static _Up* std::__copy_move<_IsMove, true, std::random_access_iterator_tag>::__copy_m(_Tp*, _Tp*, _Up*) [with _Tp = const onnxruntime::Node* const; _Up = const onnxruntime::Node*; bool _IsMove = false]',
inlined from '_OI std::__copy_move_a2(_II, _II, _OI) [with bool _IsMove = false; _II = const onnxruntime::Node* const*; _OI = const onnxruntime::Node**]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_algobase.h:506:30,
inlined from '_OI std::__copy_move_a1(_II, _II, _OI) [with bool _IsMove = false; _II = const onnxruntime::Node* const*; _OI = const onnxruntime::Node**]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_algobase.h:533:42,
inlined from '_OI std::__copy_move_a(_II, _II, _OI) [with bool _IsMove = false; _II = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, vector<const onnxruntime::Node*> >; _OI = const onnxruntime::Node**]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_algobase.h:540:31,
inlined from '_OI std::copy(_II, _II, _OI) [with _II = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, vector<const onnxruntime::Node*> >; _OI = const onnxruntime::Node**]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_algobase.h:633:7,
inlined from 'static _ForwardIterator std::__uninitialized_copy::__uninit_copy(_InputIterator, _InputIterator, _ForwardIterator) [with _InputIterator = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, std::vector<const onnxruntime::Node*> >; _ForwardIterator = const onnxruntime::Node**]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_uninitialized.h:147:27,
inlined from '_ForwardIterator std::uninitialized_copy(_InputIterator, _InputIterator, _ForwardIterator) [with _InputIterator = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, vector<const onnxruntime::Node*> >; _ForwardIterator = const onnxruntime::Node**]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_uninitialized.h:185:15,
inlined from '_ForwardIterator std::__uninitialized_copy_a(_InputIterator, _InputIterator, _ForwardIterator, allocator<_Tp>&) [with _InputIterator = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, vector<const onnxruntime::Node*> >; _ForwardIterator = const onnxruntime::Node**; _Tp = const onnxruntime::Node*]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_uninitialized.h:373:37,
inlined from 'void std::vector<_Tp, _Alloc>::_M_range_insert(iterator, _ForwardIterator, _ForwardIterator, std::forward_iterator_tag) [with _ForwardIterator = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, std::vector<const onnxruntime::Node*> >; _Tp = const onnxruntime::Node*; _Alloc = std::allocator<const onnxruntime::Node*>]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/vector.tcc:814:38,
inlined from 'std::vector<_Tp, _Alloc>::iterator std::vector<_Tp, _Alloc>::insert(const_iterator, _InputIterator, _InputIterator) [with _InputIterator = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, std::vector<const onnxruntime::Node*> >; = void; _Tp = const onnxruntime::Node*; _Alloc = std::allocator<const onnxruntime::Node*>]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_vector.h:1483:19,
inlined from 'std::vector<const onnxruntime::Node*> onnxruntime::NodeUnit::GetAllNodesInGroup() const' at /home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc:271:19:
/home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_algobase.h:437:30: error: 'void* __builtin_memmove(void*, const void*, long unsigned int)' writing between 9 and 9223372036854775800 bytes into a region of size 0 overflows the destination [-Werror=stringop-overflow=]
437 | __builtin_memmove(__result, __first, sizeof(_Tp) * _Num);
| ~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/riscv64-unknown-linux-gnu/bits/c++allocator.h:33,
from /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/allocator.h:46,
from /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/string:43:
In member function '_Tp* std::__new_allocator<_Tp>::allocate(size_type, const void*) [with _Tp = const onnxruntime::Node*]',
inlined from 'static _Tp* std::allocator_traits<std::allocator<_CharT> >::allocate(allocator_type&, size_type) [with _Tp = const onnxruntime::Node*]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/alloc_traits.h:482:28,
inlined from 'std::_Vector_base<_Tp, _Alloc>::pointer std::_Vector_base<_Tp, _Alloc>::_M_allocate(std::size_t) [with _Tp = const onnxruntime::Node*; _Alloc = std::allocator<const onnxruntime::Node*>]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_vector.h:378:33,
inlined from 'std::_Vector_base<_Tp, _Alloc>::pointer std::_Vector_base<_Tp, _Alloc>::_M_allocate(std::size_t) [with _Tp = const onnxruntime::Node*; _Alloc = std::allocator<const onnxruntime::Node*>]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_vector.h:375:7,
inlined from 'void std::vector<_Tp, _Alloc>::_M_range_insert(iterator, _ForwardIterator, _ForwardIterator, std::forward_iterator_tag) [with _ForwardIterator = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, std::vector<const onnxruntime::Node*> >; _Tp = const onnxruntime::Node*; _Alloc = std::allocator<const onnxruntime::Node*>]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/vector.tcc:805:40,
inlined from 'std::vector<_Tp, _Alloc>::iterator std::vector<_Tp, _Alloc>::insert(const_iterator, _InputIterator, _InputIterator) [with _InputIterator = __gnu_cxx::__normal_iterator<const onnxruntime::Node* const*, std::vector<const onnxruntime::Node*> >; = void; _Tp = const onnxruntime::Node*; _Alloc = std::allocator<const onnxruntime::Node*>]' at /home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/stl_vector.h:1483:19,
inlined from 'std::vector<const onnxruntime::Node*> onnxruntime::NodeUnit::GetAllNodesInGroup() const' at /home/yilyu/onnxruntime/onnxruntime/core/providers/shared/node_unit/node_unit.cc:271:19:
/home/yilyu/risc-v/riscv/riscv64-unknown-linux-gnu/include/c++/13.2.0/bits/new_allocator.h:147:55: note: at offset [-9223372036854775808, -1] into destination object of size [8, 9223372036854775800] allocated by 'operator new'
147 | return static_cast<_Tp*>(_GLIBCXX_OPERATOR_NEW(__n * sizeof(_Tp)));
| ^
cc1plus: all warnings being treated as errors

@phoebesv
Copy link
Contributor Author

@yihonglyu
This seems to be a flattbuffer issue. Please see google/flatbuffers#8119
Using the default onnxruntime compiler option, I can successfully build ORT.
I also noticed someone set -Wno-error in the ORT flatbuffer.

+ "${CMAKE_CXX_FLAGS} -Wall -pedantic -Werror -Wextra -Werror=shadow -Wno-error=stringop-overflow")

Could you try using the default ORT compile option to build it?

@phoebesv
Copy link
Contributor Author

I also found that there would be no flatbuffer issue by using config debug. I will update the documentation to use the Debug configuration until the FlatBuffers issue is resolved. I think this will provide a more stable build.

@phoebesv phoebesv force-pushed the riscv_cross_compile/gh-pages branch from a6dc816 to 9e7be4e Compare January 25, 2024 09:48
@yihonglyu
Copy link
Contributor

@yihonglyu This seems to be a flattbuffer issue. Please see google/flatbuffers#8119 Using the default onnxruntime compiler option, I can successfully build ORT. I also noticed someone set -Wno-error in the ORT flatbuffer.

+ "${CMAKE_CXX_FLAGS} -Wall -pedantic -Werror -Wextra -Werror=shadow -Wno-error=stringop-overflow")

Could you try using the default ORT compile option to build it?

That is really weird that it would work with Debug option but fail on other option. Could you possibly explain why this is happening?

@phoebesv
Copy link
Contributor Author

That is really weird that it would work with Debug option but fail on other option. Could you possibly explain why this is happening?

The compiler optimization levels for Debug and RelWithDebInfo are different, so it is possible that this problem only affects Rel.

@snnn
Copy link
Member

snnn commented Jan 27, 2024

In general we do not treat warnings from third-party libraries as errors, since there is very little thing we can do. Therefore, we may remove "-Werror" from the patch file.

@phoebesv
Copy link
Contributor Author

phoebesv commented Jan 28, 2024

In general we do not treat warnings from third-party libraries as errors, since there is very little thing we can do. Therefore, we may remove "-Werror" from the patch file.

Thanks for informing me about this. I will submit a separate pull request to address the issue with the flatbuffer.

@yihonglyu
Copy link
Contributor

@phoebesv does the PR #19238 only support the XNNPACK provider? From my understanding, it appears that it can be utilized even without XNNPACK. Could you please confirm?

@phoebesv
Copy link
Contributor Author

phoebesv commented Jan 30, 2024

@phoebesv does the PR #19238 only support the XNNPACK provider? From my understanding, it appears that it can be utilized even without XNNPACK. Could you please confirm?

xnnpack is optional, I added optional annotations in the document, hoping to provide a better understanding that xnnpack is not a mandatory option.

@phoebesv phoebesv force-pushed the riscv_cross_compile/gh-pages branch from 9e7be4e to 97f267d Compare January 30, 2024 06:08
@phoebesv phoebesv requested a review from yihonglyu January 30, 2024 06:09
@phoebesv phoebesv force-pushed the riscv_cross_compile/gh-pages branch from 97f267d to ff27131 Compare January 30, 2024 07:08
@phoebesv phoebesv force-pushed the riscv_cross_compile/gh-pages branch from ff27131 to db06b1a Compare January 30, 2024 11:38
@yihonglyu
Copy link
Contributor

yihonglyu commented Jan 30, 2024

@phoebesv, I’ve checkout out the branch at #19296 and attempted to build and test it using the command ./build.sh --parallel --config RelWithDebInfo --rv64 --riscv_toolchain_root=/home/yilyu/risc-v/riscv --riscv_qemu_path=/home/yilyu/risc-v/riscv/bin/qemu-riscv64 --build_dir /home/yilyu/onnxruntime/build/Linux/risc-v. However, numerous test cases failed. Is this an anticipated outcome? If that’s the case, do you have a subsequent PR planned to resolve the issue?

@phoebesv
Copy link
Contributor Author

@yihonglyu Our initial RISC-V PR focused on adding cross-compilation support to ORT, not all tests are enabled yet, so we've included 'skip_tests' in the build command. Considering the large number of onnx test cases and the complexities of supporting RISC-V, an initial pull request cannot possibly cover every test scenario. By enabling cross-compilation, we hope to encourage contributions to gradually enhance RISC-V support for more scenarios.

snnn pushed a commit that referenced this pull request Jan 31, 2024
### Description

Building on g++ 13.2.0 results in -Wstringop-overread errors on Linux.
This commit addresses the flatbuffer build issue with the following
changes:
1. Remove the Werror flag in the flarbuffer patch.
2. Add a compilation option to suppress the 'stringop-overflow' error in
the Flatbuffers within the xnnpack provider.

### Motivation and Context
google/flatbuffers#8119
#19239

Signed-off-by: Phoebe Chen <[email protected]>
@csukuangfj
Copy link
Contributor

Our initial RISC-V PR focused on adding cross-compilation support to ORT,

@phoebesv Thanks for supporting compiling onnxruntime for riscv64.

I have successfully compiled onnxruntime for riscv64.
However, when I use it to run my speech recognition model, the recognition results are wrong compared with using x86_64 onnxruntime.

Is that expected? Does it only support cross-compiling onnxruntime for riscv64 but the correctness has not been tested yet?

@phoebesv
Copy link
Contributor Author

phoebesv commented Feb 26, 2024

@phoebesv Thanks for supporting compiling onnxruntime for riscv64.

I have successfully compiled onnxruntime for riscv64. However, when I use it to run my speech recognition model, the recognition results are wrong compared with using x86_64 onnxruntime.

Is that expected? Does it only support cross-compiling onnxruntime for riscv64 but the correctness has not been tested yet?

@csukuangfj Yes. We currently have test scalar code by UT onnxruntime_test_all on qemu (using xnnpack EP), with multiple failed test cases. Unfortunately, we do not have enough resources to clean up all of the failed test instances.
onnxruntime_test_all riscv testing results 3688 tests from 258 test suites ran.
Failed ratio (NUM_FAILED/NUM_PASSED)~= 0.068
The most of failed test cases are due to mismatches between expected and actual results. As far as I know, a few people are working on addressing it, you are welcome to contribute.

@csukuangfj
Copy link
Contributor

using xnnpack EP

Is it possible to use the default CPU provider instead of the xnnpack provider?

@csukuangfj
Copy link
Contributor

As far as I know, a few people are working on addressing it, you are welcome to contribute.

Is there a place to find the failed tests and find whom working on which failed test?

@tianleiwu tianleiwu requested a review from yihonglyu April 5, 2024 13:50
@tianleiwu tianleiwu merged commit e31928e into microsoft:gh-pages Apr 5, 2024
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

5 participants