Eric-Song-Nop / hip-analyzer

Compiler plugin for performance analysis of HIP applications

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

HIP-analyzer

The HIP-analyzer tool provides static analysis for runtime instrumentation (tracing & software counters) for heterogeneous applications using the AMD HIP programming framework.

This tool relies on the LLVM project, the pass framework, and its C++ frontend, Clang.

Features

The current objective is to generate counters for each basic block in the compute kernel, in order to identify where divergence occurs.

Currently, the goal is to allow the user to add custom software counters and tracepoints during the execution for a fine-grained analysis of the kernel execution. We have successfully implemented various events as well as a tracing pass to this effect.

Building the tool

Using hip-analyzer requires, until further improvements, compiling ROCm-LLVM as a standalone component with BUILD_SHARED_LIBS and ENABLE_LLVM_SHARED enabled. Use the tag corresponding to your ROCm install.

Run the following command :

mkdir build && cd build
cmake -DCMAKE_C_COMPILER=/opt/rocm/llvm/bin/clang -DCMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++ -DROCM_PATH=/opt/rocm -DROCM_LLVM=<path to llvm install directory>..

Running

The HIP_CLANG_PATH environment variable must be set to the binary dir of the custom LLVM build to ensure that hipcc invokes the correct compiler.

The tool now relies on a set of LLVM IR passes which can be found in the libhip-analyzer-pass.so shared library. It can be run using the opt tool on LLVM IR modules, or directly in the hipcc compiler, using a few compiler flags.

hipcc (prefered way, one-shot compilation):

hipcc <input> -o <output> -fpass-plugin=libhip-analyzer-passes.so

opt :

opt -load=libhip-analyzer-passes.so <module> [--hip-analyzer-counters | --hip-analyzer-trace | --hip-analyzer-host]

The flags to use depends on the module you're using (whether it be host or device)

About

Compiler plugin for performance analysis of HIP applications

License:Apache License 2.0


Languages

Language:C++ 90.0%Language:CMake 5.7%Language:C 4.3%