Merge pull request #4995 from rongou/per-thread-default-stream

[REVIEW] Add CMake option for per-thread default stream

CHANGELOG.md

@@ -21,6 +21,7 @@
 
 ## Improvements
 
+- PR #4995 Add CMake option for per-thread default stream
 - PR #5033 Fix Numba deprecations warnings with Numba 0.49+
 - PR #4950 Fix import errors with Numba 0.49+
 - PR #4825 Update the iloc exp in dataframe.py

cpp/CMakeLists.txt

@@ -275,6 +275,16 @@ if(JITIFY_USE_CACHE)
     add_definitions("-DJITIFY_USE_CACHE -DCUDF_VERSION=${CMAKE_PROJECT_VERSION}")
 endif(JITIFY_USE_CACHE)
 
+###################################################################################################
+# - per-thread default stream option --------------------------------------------------------------
+# This needs to be defined first so tests and benchmarks can inherit it.
+
+option(PER_THREAD_DEFAULT_STREAM "Build with per-thread default stream" OFF)
+if(PER_THREAD_DEFAULT_STREAM)
+    message(STATUS "Using per-thread default stream")
+    add_compile_definitions(CUDA_API_PER_THREAD_DEFAULT_STREAM)
+endif(PER_THREAD_DEFAULT_STREAM)
+
 ###################################################################################################
 # - add gtest -------------------------------------------------------------------------------------
 

java/README.md

@@ -77,3 +77,34 @@ You will get errors if you don't do it consistently.  We tried to detect these u
 
 If you have a compatible GPU on your build system the tests will use it.  If not you will see a
 lot of skipped tests.
+
+## Per-thread Default Stream
+
+The JNI code can be built with *per-thread default stream* (PTDS), which gives each host thread its
+own default CUDA stream, and can potentially increase the overlap of data copying and compute
+between different threads (see
+[blog post](https://devblogs.nvidia.com/gpu-pro-tip-cuda-7-streams-simplify-concurrency/)).
+
+Since the PTDS option is for each compilation unit, it should be done at the same time across the
+whole codebase. To enable PTDS, first build RMM:
+```shell script
+conda activate cudf_dev
+cd src/rmm/build
+cmake .. -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX -DPER_THREAD_DEFAULT_STREAM=ON
+make -j`nproc`
+make install
+```
+
+then build cuDF:
+```shell script
+cd src/cudf/cpp/build
+cmake .. -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX -DPER_THREAD_DEFAULT_STREAM=ON
+make -j`nproc`
+make install
+```
+
+and finally build the jar:
+```shell script
+cd src/cudf/java
+mvn clean install -DPER_THREAD_DEFAULT_STREAM=ON
+```

java/pom.xml

@@ -144,6 +144,7 @@
         <skipNativeCopy>false</skipNativeCopy>
         <cxx11.abi.value>ON</cxx11.abi.value>
         <CUDA_STATIC_RUNTIME>OFF</CUDA_STATIC_RUNTIME>
+        <PER_THREAD_DEFAULT_STREAM>OFF</PER_THREAD_DEFAULT_STREAM>
     </properties>
 
     <profiles>
@@ -333,6 +334,7 @@
                                       executable="cmake">
                                     <arg value="${basedir}/src/main/native"/>
                                     <arg value="-DCUDA_STATIC_RUNTIME=${CUDA_STATIC_RUNTIME}" />
+                                    <arg value="-DPER_THREAD_DEFAULT_STREAM=${PER_THREAD_DEFAULT_STREAM}" />
                                     <arg value="-DCMAKE_CXX11_ABI=${cxx11.abi.value}"/>
                                 </exec>
                                 <exec dir="${native.build.path}"

java/src/main/native/CMakeLists.txt

@@ -234,6 +234,12 @@ if(USE_NVTX)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_NVTX")
 endif(USE_NVTX)
 
+option(PER_THREAD_DEFAULT_STREAM "Build with per-thread default stream" OFF)
+if(PER_THREAD_DEFAULT_STREAM)
+    message(STATUS "Using per-thread default stream")
+    add_compile_definitions(CUDA_API_PER_THREAD_DEFAULT_STREAM)
+endif(PER_THREAD_DEFAULT_STREAM)
+
 ###################################################################################################
 # - link libraries --------------------------------------------------------------------------------