Fix segfault in release build with GCC 5. (#419)

* Use GCC 6 and GCC 7 in GitHub actions to prevent segfault.

GCC 5 will cause segfault in Release build.

* Switch to GCC 5.

Disable inline optimization of `CudaStreamOverride`.

* Move Push and Pop of CudaStreamOverride to context.cu

It prevents the compiler from inlining them.

.github/workflows/build.yml

@@ -24,7 +24,7 @@ jobs:
       matrix:
         os: [ubuntu-16.04, ubuntu-18.04]
         cuda: ["10.1", "10.2"]
-        gcc: ["5", "6"]
+        gcc: ["5"]
         torch: ["1.6.0", "1.7.0"]
         python-version: [3.6, 3.7, 3.8]
 

.github/workflows/run-tests.yml

@@ -20,15 +20,11 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-18.04]
-        cuda: ["10.1", "10.2"]
+        cuda: ["10.1"]
         gcc: ["5"]
-        torch: ["1.6.0", "1.7.0"]
+        torch: ["1.6.0"]
         python-version: [3.6]
-        # build_type: ["Release", "Debug"]
-        #
-        # disable release build for github actions since it results
-        # in segfault which CANNOT be reproduced locally.
-        build_type: ["Debug"]
+        build_type: ["Release", "Debug"]
 
     steps:
       # refer to https://github.com/actions/checkout
@@ -56,6 +52,10 @@ jobs:
           echo "CXX=/usr/bin/g++-${{ matrix.gcc }}" >> $GITHUB_ENV
           echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}" >> $GITHUB_ENV
 
+      - name: Display GCC version
+        run: |
+          gcc --version
+
       - name: Setup Python ${{ matrix.python-version }}
         uses: actions/setup-python@v2
         with:
@@ -70,11 +70,10 @@ jobs:
           torch: ${{ matrix.torch }}
         shell: bash
         run: |
-          python3 -m pip install --upgrade pip
-          python3 -m pip install wheel twine
-          python3 -m pip install bs4 requests tqdm
-          python3 -m pip install dataclasses graphviz
-          sudo apt-get install python-pydot python-pydot-ng graphviz
+          python3 -m pip install -qq --upgrade pip six
+          python3 -m pip install -qq bs4 requests tqdm
+          python3 -m pip install -qq dataclasses graphviz
+          sudo apt-get -qq install graphviz
 
           ./scripts/github_actions/install_torch.sh
           python3 -c "import torch; print('torch version:', torch.__version__)"
@@ -94,12 +93,23 @@ jobs:
           cmake -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ..
           cat k2/csrc/version.h
 
-      - name: Build and Run Tests
+      - name: ${{ matrix.build_type }} Build
         shell: bash
         run: |
           echo "number of cores: $(nproc)"
           cd build
           # we cannot use -j here because of limited RAM
           # of the VM provided by GitHub actions
           make
+
+      - name: Display Build Information
+        shell: bash
+        run: |
+          export PYTHONPATH=$PWD/k2/python:$PWD/build/lib:$PYTHONPATH
+          python3 -m k2.version
+
+      - name: Run Tests
+        shell: bash
+        run: |
+          cd build
           ctest --output-on-failure

.github/workflows/style_check.yml

@@ -14,10 +14,10 @@ on:
 
 jobs:
   build:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-18.04
     strategy:
       matrix:
-        python-version: [3.7, 3.8]
+        python-version: [3.7]
 
     steps:
       - uses: actions/checkout@v2

INSTALL.md

@@ -41,7 +41,7 @@ you still need to install PyTorch with CUDA support.
 
 There are two ways to install k2 from pre-built wheel packages.
 
-### (1) From PyPI using `pip install k2`
+### (1) From PyPI using `pip install --pre k2`
 
   The wheel packages on PyPI are built using torch==1.6.0+cu101 on Ubuntu 18.04.
   If you are using other Linux systems, the pre-built wheel packages may NOT
@@ -150,10 +150,13 @@ make -j
 ctest --parallel <JOBNUM>
 ```
 
-If Valgrind is installed, you can check heap corruptions and memory leaks by
+If `valgrind` is installed, you can check heap corruptions and memory leaks by
 
 ```bash
 cd build
 make -j
 ctest -R <TESTNAME> -D ExperimentalMemCheck
 ```
+
+**HINT**: You can install `valgrind` with `sudo apt-get install valgrind`
+on Ubuntu.

k2/csrc/context.cu

@@ -15,6 +15,17 @@
 
 namespace k2 {
 
+void CudaStreamOverride::Push(cudaStream_t stream) {
+  stack_.push_back(stream);
+  stream_override_ = stream;
+}
+
+void CudaStreamOverride::Pop(cudaStream_t stream) {
+  K2_DCHECK(!stack_.empty());
+  K2_DCHECK_EQ(stack_.back(), stream);
+  stack_.pop_back();
+}
+
 RegionPtr NewRegion(ContextPtr context, std::size_t num_bytes) {
   // .. fairly straightforward.  Sets bytes_used to num_bytes, caller can
   // overwrite if needed.

k2/csrc/context.h

@@ -374,18 +374,14 @@ class CudaStreamOverride {
     else
       return stream;
   }
-  void Push(cudaStream_t stream) {
-    stack_.push_back(stream);
-    stream_override_ = stream;
-  }
-  void Pop(cudaStream_t stream) {
-    K2_DCHECK(!stack_.empty());
-    K2_DCHECK_EQ(stack_.back(), stream);
-    stack_.pop_back();
-  }
+
+  void Push(cudaStream_t stream);
+
+  void Pop(cudaStream_t stream);
 
   CudaStreamOverride() : stream_override_(0x0) {}
 
+ private:
   cudaStream_t stream_override_;
   std::vector<cudaStream_t> stack_;
 };

scripts/github_actions/install_cuda.sh

@@ -10,8 +10,11 @@ case "$cuda" in
     filename=cuda_10.0.130_410.48_linux
     ;;
   10.1)
-    url=https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.105_418.39_linux.run
-    filename=cuda_10.1.105_418.39_linux.run
+    # WARNING: there are bugs in
+    # https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.105_418.39_linux.run
+    # with GCC 7. Please use the following version
+    url=http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.243_418.87.00_linux.run
+    filename=cuda_10.1.243_418.87.00_linux.run
     ;;
   10.2)
     url=http://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run