Vendor sleef as a submodule (#130)

.github/workflows/build-test.yml

@@ -56,6 +56,8 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
+        with:
+          submodules: recursive
 
       - name: Install Python ${{ matrix.python }}
         uses: actions/setup-python@v5

.gitignore

@@ -9,6 +9,7 @@ python/triton*.egg-info/
 
 python/triton/_C/*.pyd
 python/triton/_C/*.so
+python/triton/_C/*.so.*
 python/triton/_C/*.dylib
 python/triton/_C/*.pdb
 python/triton/_C/*.exe

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "sleef"]
+	path = third_party/sleef
+	url = https://github.com/shibatch/sleef

third_party/cpu/CMakeLists.txt

@@ -5,5 +5,16 @@ add_subdirectory(lib)
 if(TRITON_BUILD_PYTHON_MODULE)
   add_triton_plugin(TritonCPU ${CMAKE_CURRENT_SOURCE_DIR}/triton_cpu.cc LINK_LIBS TritonCPUToLLVM TritonCPUTransforms)
   target_link_libraries(TritonCPU PUBLIC MLIRVectorToSCF MLIRAffineToStandard MLIRMathToLibm)
-  add_library(TritonCPURuntime SHARED ${CMAKE_CURRENT_SOURCE_DIR}/runtime/cpu_runtime.cpp)
 endif()
+
+add_library(TritonCPURuntime SHARED ${CMAKE_CURRENT_SOURCE_DIR}/runtime/cpu_runtime.cpp)
+
+# Build and link sleef
+set(SLEEF_BUILD_SHARED_LIBS ON CACHE BOOL "Build sleef shared lib" FORCE)
+set(SLEEF_BUILD_DFT OFF CACHE BOOL "Don't build sleef DFT lib" FORCE)
+set(SLEEF_BUILD_GNUABI_LIBS OFF CACHE BOOL "Don't build sleef gnuabi libs" FORCE)
+set(SLEEF_BUILD_TESTS OFF CACHE BOOL "Don't build sleef tests" FORCE)
+set(SLEEF_BUILD_SCALAR_LIB OFF CACHE BOOL "libsleefscalar will not be built." FORCE)
+add_subdirectory("${CMAKE_SOURCE_DIR}/third_party/sleef" sleef)
+# Override sleef's output directory with our own
+set_target_properties(sleef PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_LIBRARY_OUTPUT_DIRECTORY})

third_party/cpu/backend/compiler.py

@@ -215,14 +215,7 @@ def make_so(src, metadata, options):
             asm_path = os.path.join(tmpdir, "kernel.s")
             Path(asm_path).write_text(src)
             lib_dirs = cpu_driver.library_dirs
-            libs = ["gcc", "m", "TritonCPURuntime"]
-            # TRITON_CPU_USE_SLEEF=1 - use system libsleef
-            # TRITON_CPU_USE_SLEEF=path - use libsleef from the specified path
-            use_sleef = os.environ.get("TRITON_CPU_USE_SLEEF", "0")
-            if use_sleef != "0":
-                if os.path.isdir(use_sleef):
-                    lib_dirs.append(use_sleef)
-                libs.append("sleef")
+            libs = ["gcc", "m", "TritonCPURuntime", "sleef"]
             so = _build("kernel", asm_path, tmpdir, lib_dirs, cpu_driver.include_dirs, libs)
             with open(so, "rb") as f:
                 return f.read()