[CI] Made build-wheels compatible with system LLVM setup (#138)

This speeds up wheelhouse build time by ~10x

.ci/build-wheels.yml

@@ -16,7 +16,7 @@ jobs:
         set -o errexit
         python3 --version
         python3 -m pip install --upgrade pip
-        pip3 install cibuildwheel==1.10.0
+        pip3 install cibuildwheel==2.0.0
         pip3 install twine
       displayName: Install dependencies
     - bash: |
@@ -28,7 +28,11 @@ jobs:
         echo "base-dir=/project" >> python/setup.cfg
       displayName: Patch setup.py
     - bash: |
-        export CIBW_BEFORE_BUILD="pip install cmake"
+        export CIBW_MANYLINUX_X86_64_IMAGE="manylinux2014"
+        export CIBW_MANYLINUX_PYPY_X86_64_IMAGE="manylinux2014"
+        export CIBW_BEFORE_BUILD="pip install cmake;\
+                                  yum install -y llvm11 llvm11-devel llvm11-static llvm11-libs zlib-devel;"
+        export CIBW_SKIP="{cp,pp}35-*"
         export CIBW_BUILD="{cp,pp}3*-manylinux_x86_64"
         python3 -m cibuildwheel python --output-dir wheelhouse
       displayName: Build wheels

CMakeLists.txt

@@ -53,7 +53,6 @@ if(BUILD_PYTHON_MODULE)
         add_definitions(-DWITH_CUTLASS_BINDINGS)
         set(CUTLASS_LIBRARIES "cutlass.a")
     endif()
-    message(STATUS ${CUTLASS_INCLUDE_PATH})
     include_directories("." ${PYTHON_SRC_PATH} ${PYTHON_INCLUDE_DIRS} ${CUTLASS_INCLUDE_DIR})
     link_directories(${PYTHON_LINK_DIRS} ${CUTLASS_LIBRARY_DIR})
     set(PYTHON_SRC ${PYTHON_SRC_PATH}/main.cc ${PYTHON_SRC_PATH}/triton.cc  ${PYTHON_SRC_PATH}/superblock.cc ${CUTLASS_SRC})
@@ -65,7 +64,7 @@ file(GLOB_RECURSE LIBTRITON_SRC lib/*.cc)
 add_library(triton SHARED ${LIBTRITON_SRC} ${PYTHON_SRC})
 target_link_options(triton PRIVATE ${LLVM_LDFLAGS})
 target_link_libraries(triton ${LLVM_LIBRARIES} ${LLVM_SYSTEM_LIBS})
-message(STATUS ${LLVM_LDFLAGS})
+
 
 if(BUILD_PYTHON_MODULE)
     set(CMAKE_SHARED_LIBRARY_SUFFIX ".so")

cmake/FindLLVM.cmake

@@ -33,12 +33,12 @@
 # We also want an user-specified LLVM_ROOT_DIR to take precedence over the
 # system default locations such as /usr/local/bin. Executing find_program()
 # multiples times is the approach recommended in the docs.
-set(llvm_config_names llvm-config-12.0 llvm-config120 llvm-config-12
+set(llvm_config_names llvm-config-12.0 llvm-config120 llvm-config-12 llvm-config-12-64
-                      llvm-config-11.0 llvm-config110 llvm-config-11
+                      llvm-config-11.0 llvm-config110 llvm-config-11 llvm-config-11-64
-                      llvm-config-10.0 llvm-config100 llvm-config-10
+                      llvm-config-10.0 llvm-config100 llvm-config-10 llvm-config-10-64
-                      llvm-config-9.0 llvm-config90 llvm-config-9
+                      llvm-config-9.0 llvm-config90 llvm-config-9 llvm-config-9-64
-                      llvm-config-8.0 llvm-config80 llvm-config-8
+                      llvm-config-8.0 llvm-config80 llvm-config-8 llvm-config-8-64
-                      llvm-config-7.0 llvm-config70 llvm-config-7
+                      llvm-config-7.0 llvm-config70 llvm-config-7 llvm-config-7-64
                       llvm-config-6.0 llvm-config60
                       llvm-config)
 find_program(LLVM_CONFIG

include/triton/ir/dispatch.h

@@ -4,6 +4,7 @@
 #define _TRITON_IR_DISPATCH_H_
 
 #include "triton/ir/builder.h"
+#include <stdexcept>
 
 namespace triton{
 namespace ir{

lib/ir/constant.cc

@@ -1,4 +1,5 @@
 #include <cassert>
+#include <stdexcept>
 #include "triton/ir/constant.h"
 #include "triton/ir/type.h"
 #include "triton/ir/context.h"

lib/ir/type.cc

@@ -1,4 +1,5 @@
 #include <cassert>
+#include <stdexcept>
 #include "triton/ir/type.h"
 #include "triton/ir/context.h"
 #include "triton/ir/context_impl.h"

python/setup.py

@@ -7,12 +7,14 @@ import subprocess
 import distutils
 import glob
 import tempfile
+import shutil
 from distutils.version import LooseVersion
 from setuptools import setup, Extension, find_packages
 from setuptools.command.build_ext import build_ext
 from setuptools.command.test import test as TestCommand
 import distutils.spawn
-
+import urllib.request
+import tarfile
 
 class CMakeExtension(Extension):
     def __init__(self, name, path, sourcedir=""):
@@ -66,7 +68,7 @@ class CMakeBuild(build_ext):
             "-DBUILD_TUTORIALS=OFF",
             "-DBUILD_PYTHON_MODULE=ON",
             #'-DPYTHON_EXECUTABLE=' + sys.executable,
-            '-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON',
+            #'-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON',
             "-DTRITON_LLVM_BUILD_DIR=" + llvm_build_dir,
             "-DPYTHON_INCLUDE_DIRS=" + ";".join(python_include_dirs)
         ]