NVIDIA Looking To Create New Tool For Generating AutoFDO Profiles For GCC
NVIDIA compiler engineers are looking to develop a standalone tool that could be upstreamed into the GNU Compiler Collection (GCC) codebase for generating AutoFDO profiles for consumption by GCC in turn for better benefiting from automatic feedback directed optimizations (FDO) in the name of better performance.
Due to the current AutoFDO tools from Google not being used by GCC but rather tailored for LLVM/Clang, NVIDIA engineers are stepping up in pursuing an AutoFDO profile creation tool for GCC. This tool would be catered to GCC, aims to be lightweight and memory efficient, and work well for creating AutoFDO profiles.
Automatic Feedback Directed Optimizations can be of significant performance benefit when engaged by the compiler, but relies on sampling profiles of the program being built in order for the compiler to make more accurate assessments for optimizations. With this proposed NVIDIA tool it would help developers in creating those AutoFDO profiles for consumption by GCC during the secondary builds for leveraging that profile data. This tool would be catered to GCC, aims to be lightweight and memory efficient, and work well for creating AutoFDO profiles in pairing with the GNU toolchain.
Google uses AutoFDO on the likes of Android and previously have shown off some very strong benefits but critical is collecting the necessary sample profiles:
The request for comments on this proposed tool was sent out to the GCC mailing list by NVIDIA engineer Kugan Vivekanandarajah. See this GCC mailing list thread for details on the proposal. It's great seeing NVIDIA stepping up to this open-source contribution for addressing this current optimization shortcoming relative to LLVM/Clang's AutoFDO support.
Due to the current AutoFDO tools from Google not being used by GCC but rather tailored for LLVM/Clang, NVIDIA engineers are stepping up in pursuing an AutoFDO profile creation tool for GCC. This tool would be catered to GCC, aims to be lightweight and memory efficient, and work well for creating AutoFDO profiles.
Automatic Feedback Directed Optimizations can be of significant performance benefit when engaged by the compiler, but relies on sampling profiles of the program being built in order for the compiler to make more accurate assessments for optimizations. With this proposed NVIDIA tool it would help developers in creating those AutoFDO profiles for consumption by GCC during the secondary builds for leveraging that profile data. This tool would be catered to GCC, aims to be lightweight and memory efficient, and work well for creating AutoFDO profiles in pairing with the GNU toolchain.
Google uses AutoFDO on the likes of Android and previously have shown off some very strong benefits but critical is collecting the necessary sample profiles:
The request for comments on this proposed tool was sent out to the GCC mailing list by NVIDIA engineer Kugan Vivekanandarajah. See this GCC mailing list thread for details on the proposal. It's great seeing NVIDIA stepping up to this open-source contribution for addressing this current optimization shortcoming relative to LLVM/Clang's AutoFDO support.