Enabling c++20 on linux #17816
Enabling c++20 on linux #17816
Conversation
…ERTY LINK_FLAGS " -s ASSERTIONS=2 -s SAFE_HEAP=1 -s STACK_OVERFLOW_CHECK=2")"
…ING PROPERTY LINK_FLAGS " -s ASSERTIONS=2 -s SAFE_HEAP=1 -s STACK_OVERFLOW_CHECK=2")"" This reverts commit c72c40e.
|
We cannot do it for Linux CUDA build yet. Because the compiler we use is GCC 8, which is too low. |
|
Just out of curiosity. Why do you need C++20? Would it crash onnxruntime build in my old conda environment? |
|
For your second question, no. A new C++ standard consists of two things:
The first one doesn't have impact on compatibility. The second one has impacts on non-Windows/non-Linux systems like macOS/iOS. We can avoid the issue by not using the headers that are not supported by the target system. |
|
Wait this #20786 . The PR will update all GCCs to >=11. Now we still have GCC 8. |
|
The GCC version is updated. |
We can use the new C++ features to write better code, and it shouldn't impact our compatibility at runtime. The language features are listed in https://gcc.gnu.org/projects/cxx-status.html in "C++20 Language Features" section |
| @@ -326,11 +326,11 @@ class TreeAggregatorMin : public TreeAggregator<InputType, ThresholdType, Output | |||
| template <typename InputType, typename ThresholdType, typename OutputType> | |||
| class TreeAggregatorMax : public TreeAggregator<InputType, ThresholdType, OutputType> { | |||
| public: | |||
| TreeAggregatorMax<InputType, ThresholdType, OutputType>(size_t n_trees, | |||
There was a problem hiding this comment.
All other similar class like TreeAggregatorMin and TreeAggregatorSum does not include type qualifier.
|
depends on #21071 |
…nux_c++20 # Conflicts: # onnxruntime/contrib_ops/cuda/math/cufft_plan_cache.h # orttraining/orttraining/training_ops/cuda/nn/conv_shared.h # orttraining/orttraining/training_ops/rocm/nn/conv_grad.cc
for (size_t x = 0; x < body; x++) {
volatile_x = x;
// Your code here, using volatile_x if necessary
}
#pragma GCC diagnostic ignored "-Wdeprecated"
#pragma GCC diagnostic ignored "-Wdeprecated"
#pragma clang diagnostic pop #endif
# Conflicts: # onnxruntime/core/providers/cpu/ml/tree_ensemble_aggregator.h
| @@ -108,7 +108,14 @@ std::cv_status OrtCondVar::wait_for(std::unique_lock<OrtMutex>& cond_mutex, | |||
| namespace onnxruntime { | |||
|
|
|||
| class OrtMutex { | |||
| #if defined(__clang__) && __cplusplus >= 202002L | |||
| #pragma clang diagnostic push | |||
| #pragma clang diagnostic ignored "-Wdeprecated-pragma" | |||
There was a problem hiding this comment.
I think this can be fixed - see this patch: e34c91c
| @@ -87,7 +87,18 @@ Status EmbedLayerNorm<T>::Compute(OpKernelContext* context) const { | |||
|
|
|||
| int n = batch_size * sequence_length; | |||
| concurrency::ThreadPool::TryBatchParallelFor( | |||
| #if __cplusplus >= 202002L | |||
| context->GetOperatorThreadPool(), n, [=, this, &failed](ptrdiff_t index) { | |||
There was a problem hiding this comment.
see this core guideline:
https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines#f54-when-writing-a-lambda-that-captures-this-or-any-class-data-member-dont-use--default-capture
I think they have a point. at least, we should consider how to make the code easier to understand.
|
Now we are using NDK 27? |
Description
Enabling c++20 on Linux
Currently blocking issue:
Motivation and Context
We want the latest and the greatest features from c++20.
depends on