diff --git a/profiler/README.md b/profiler/README.md index 05bbc7b4f9..86f668eacb 100644 --- a/profiler/README.md +++ b/profiler/README.md @@ -1,5 +1,23 @@ [Back to the main page](../README.md) # Composable Kernel profiler + +## Building Specific Profilers +To reduce build time, filter which operations to compile using CMake options: + +```bash +# Build all grouped_gemm variants (grouped_gemm, grouped_gemm_fastgelu, grouped_gemm_tile_loop, etc.) +cmake -DCK_PROFILER_OP_FILTER="grouped_gemm" .. + +# Build ONLY base grouped_gemm (excludes variants - use exact regex match with ^ and $) +cmake -DCK_PROFILER_OP_FILTER="^grouped_gemm$" .. +``` + +Both `CK_PROFILER_OP_FILTER` and `CK_PROFILER_INSTANCE_FILTER` accept regex patterns. Default builds all operations. + +To find the complete list of operations, run the following command: +```bash +find profiler/src -name "profile_*.cpp" | sed 's|profiler/src/profile_||' | sed 's|.cpp||' | sort +``` ## Profiler GEMM UNIVERSAL kernels ```bash # arg1: tensor operation (gemm_universal: Universal GEMM)