Revert "[reland][inductor] Add an AOT compilation mode for Inductor CPP backend (#95985)"

This reverts commit deaf9e5.

Reverted #95985 on behalf of https://github.com/huydhn due to Sorry for reverting this. It increased the test time significantly for ASAN (and may be other test shards). ASAN tests on PR passed but it was barely not timing out. I have updated my initial findings in #96378

Author: pytorchmergebot

test/inductor/aot/cpp/CMakeLists.txt

@@ -27,27 +27,6 @@ def compile(
     return compile_fx(gm, example_inputs, config_patches=options)
 
 
-def aot_compile(
-    gm: torch.fx.GraphModule,
-    example_inputs: List[torch.Tensor],
-    options: Optional[Dict[str, Any]] = None,
-) -> str:
-    """
-    Ahead-of-time compile a given FX graph with TorchInductor into a shared library.
-
-    Args:
-        gm: The FX graph to compile.
-        example_inputs:  List of tensor inputs.
-        options:  Optional dict of config options.  See `torch._inductor.config`.
-
-    Returns:
-        Path to the generated shared library
-    """
-    from .compile_fx import compile_fx
-
-    return compile_fx(gm, example_inputs, config_patches=options, aot_mode=True)()
-
-
 def list_mode_options(mode: str = None) -> Dict[str, Any]:
     r"""Returns a dictionary describing the optimizations that each of the available
     modes passed to `torch.compile()` performs.

test/inductor/aot/cpp/test.cpp

@@ -534,52 +534,6 @@ def cpp_compile_command(
     ).strip()
 
 
-class AotCodeCache:
-    cache = dict()
-    clear = staticmethod(cache.clear)
-
-    @classmethod
-    def compile(cls, source_code):
-        from .codegen.wrapper import CppWrapperCodeGen
-
-        # TODO: update cpp_compile_command for different platforms
-        picked_vec_isa = pick_vec_isa()
-        key, input_path = write(
-            source_code,
-            "cpp",
-            code_hash(repr(cpp_compile_command("i", "o", vec_isa=picked_vec_isa))),
-        )
-        if key not in cls.cache:
-            from filelock import FileLock
-
-            lock_dir = get_lock_dir()
-            lock = FileLock(os.path.join(lock_dir, key + ".lock"), timeout=LOCK_TIMEOUT)
-            with lock:
-                output_so = (
-                    os.path.join(os.getcwd(), f"{config.aot_codegen_output_prefix}.so")
-                    if config.aot_codegen_output_prefix
-                    else f"{input_path[:-3]}.so"
-                )
-
-                output_header = f"{output_so[:-3]}.h"
-                with open(output_header, "w") as header_file:
-                    header_file.writelines("#include <torch/torch.h>\n\n")
-                    header_file.writelines(f"{CppWrapperCodeGen.decl_str};\n")
-
-                log.info(f"AOT-Inductor compiles code into: {output_so}")
-                if not os.path.exists(output_so):
-                    cmd = cpp_compile_command(
-                        input=input_path, output=output_so, vec_isa=picked_vec_isa
-                    ).split(" ")
-                    try:
-                        subprocess.check_output(cmd, stderr=subprocess.STDOUT)
-                    except subprocess.CalledProcessError as e:
-                        raise exc.CppCompileError(cmd, e.output) from e
-
-                cls.cache[key] = output_so
-        return cls.cache[key]
-
-
 class CppCodeCache:
     cache = dict()
     clear = staticmethod(cache.clear)

test/inductor/aot/cpp/test.py

@@ -2050,12 +2050,7 @@ def codegen_define_and_call(self, wrapper):
         )
         if enable_kernel_profile:
             code.writelines(["#include <ATen/record_function.h>"])
-        kernel_decl_name = kernel_name if V.graph.aot_mode else "kernel"
-
-        if not V.graph.aot_mode or self.count == 1:
-            code.writeline(cpp_prefix())
-
-        code.writeline(f'extern "C" void {kernel_decl_name}({arg_defs})')
+        code.writelines([cpp_prefix(), "" f'extern "C" void kernel({arg_defs})'])
         with code.indent():
             if enable_kernel_profile:
                 graph_id = V.graph.graph_id
@@ -2070,12 +2065,9 @@ def codegen_define_and_call(self, wrapper):
             code.splice(self.loops_code)
 
         codecache_def = IndentedBuffer()
-        if V.graph.aot_mode:
-            codecache_def.splice(code)
-        else:
-            codecache_def.writeline("async_compile.cpp('''")
-            codecache_def.splice(code)
-            codecache_def.writeline("''')")
+        codecache_def.writeline("async_compile.cpp('''")
+        codecache_def.splice(code)
+        codecache_def.writeline("''')")
 
         codecache_str = codecache_def.getvalue()
         # TODO(voz): Ostensibly, we should not need this. But there are cases where C++ codegen does

test/inductor/aot/cpp/test.sh

@@ -5,7 +5,6 @@
 #include <limits>
 #include <omp.h>
 
-#include <ATen/ATen.h>
 #include <ATen/core/PhiloxRNGEngine.h>
 #if defined(CPU_CAPABILITY_AVX512) || defined(CPU_CAPABILITY_AVX2)
 #include <ATen/cpu/vec/functional.h>