[BACKEND][CPU] Make it buildable and runnable in a different environm…

…ent (triton-lang#8)

* [BACKEND][CPU] Make it buildable and runnable in a different environment

* Revert seemingly inconsistent python code formatting

include/triton/Conversion/CMakeLists.txt

@@ -1,4 +1,5 @@
-add_subdirectory(TritonCPUToLLVM)
+# TODO(minjang): I will remove these scratches soon.
+# add_subdirectory(TritonCPUToLLVM)
 add_subdirectory(TritonGPUToLLVM)
-add_subdirectory(TritonToTritonCPU)
+# add_subdirectory(TritonToTritonCPU)
 add_subdirectory(TritonToTritonGPU)

lib/Conversion/CMakeLists.txt

@@ -1,4 +1,5 @@
-#add_subdirectory(TritonToTritonCPU)
+# TODO(minjang): I will remove these scratches soon.
+# add_subdirectory(TritonToTritonCPU)
 add_subdirectory(TritonToTritonGPU)
-#add_subdirectory(TritonCPUToLLVM)
+# add_subdirectory(TritonCPUToLLVM)
 add_subdirectory(TritonGPUToLLVM)

python/src/passes.cc

@@ -6,7 +6,6 @@
 #include "triton/Analysis/Allocation.h"
 #include "triton/Analysis/Membar.h"
 #include "triton/Conversion/TritonGPUToLLVM/Passes.h"
-#include "triton/Conversion/TritonToTritonCPU/Passes.h"
 #include "triton/Conversion/TritonToTritonGPU/Passes.h"
 #include "triton/Dialect/Triton/Transforms/Passes.h"
 #include "triton/Dialect/TritonGPU/Transforms/Passes.h"
@@ -44,8 +43,6 @@ void init_triton_passes_ttir(py::module &&m) {
   ADD_PASS_WRAPPER_4("add_convert_to_ttgpuir",
                      createConvertTritonToTritonGPUPass, const std::string &,
                      int, int, int);
-  // ADD_PASS_WRAPPER_0("add_convert_to_ttcpuir",
-  //                    createConvertTritonToTritonCPUPass);
 }
 
 void init_triton_passes_ttgpuir(py::module &&m) {

python/triton/runtime/build.py

@@ -47,6 +47,9 @@ def _build(name, src, srcdir, library_dirs, include_dirs, libraries):
     cc_cmd += [f'-l{lib}' for lib in libraries]
     cc_cmd += [f"-L{dir}" for dir in library_dirs]
     cc_cmd += [f"-I{dir}" for dir in include_dirs if dir is not None]
+    # CPU backend uses C++ (driver.cpp). Some old version compilers need a specific C++17 flag.
+    if src.endswith(".cpp") or src.endswith(".cc"):
+        cc_cmd += ["-std=c++17"]
     ret = subprocess.check_call(cc_cmd)
     if ret == 0:
         return so

third_party/cpu/backend/compiler.py

@@ -20,7 +20,7 @@ class CPUOptions:
     cluster_dims: tuple = (1, 1, 1)
     extern_libs: dict = None
     debug: bool = False
-    allowed_dot_input_precisions: Tuple[str] = ("ieee",)
+    allowed_dot_input_precisions: Tuple[str] = ("ieee", )
     allow_fp8e4nv: bool = False
     enable_fp_fusion: bool = True
 

third_party/cpu/backend/driver.py

@@ -74,6 +74,7 @@
     "LLVMSupport",
     "LLVMDemangle",
     "stdc++",
+    "z",
 ]
 
 
@@ -176,7 +177,8 @@ def format_of(ty):
     arg_ptrs_list = ', '.join(f"&arg{i}" for i, ty in signature.items()) if len(signature) > 0 else ''
     kernel_fn_args = [i for i in signature.keys() if i not in constants]
     kernel_fn_args_list = ', '.join(f"arg{i}" for i in kernel_fn_args) if len(kernel_fn_args) > 0 else ''
-    kernel_fn_arg_types = (', '.join(f"{ty_to_cpp(signature[i])}" for i in kernel_fn_args) + ", " if len(signature) > 0 else '') + "uint32_t, uint32_t, uint32_t"
+    kernel_fn_arg_types = (', '.join(f"{ty_to_cpp(signature[i])}" for i in kernel_fn_args) +
+                           ", " if len(signature) > 0 else '') + "uint32_t, uint32_t, uint32_t"
 
     # generate glue code
     src = f"""

third_party/cpu/include/TritonCPUToLLVM/CMakeLists.txt

@@ -1,3 +1,3 @@
 set(LLVM_TARGET_DEFINITIONS Passes.td)
 mlir_tablegen(Passes.h.inc -gen-pass-decls --name TritonCPUToLLVM)
-add_public_tablegen_target(TritonCPUConversionPassIncGen)
+add_public_tablegen_target(TritonCPUToLLVMConversionPassIncGen)