Dtype selective build: enable in fbcode

^
By removing `header_namespace` and depending on #include "selected_op_variants.h" instead of the full path #include <executorch/kernels/portable/cpu/selected_op_variants.h>

Note:
expose_operator_symbols=False only works in xplat, so add a failure message for that too. I don't think we should recommend for users to set it to true, as it prevents a library from linking multiple executorch_generated_libs (symbols will clash).

Differential Revision: [D75082395](https://our.internmc.facebook.com/intern/diff/D75082395/)

[ghstack-poisoned]

Author: lucylq

kernels/portable/cpu/pattern/targets.bzl

@@ -11,7 +11,7 @@ def define_common_targets():
     # build, where the portable ops are built from source and linked with :all_deps
     runtime.cxx_library(
         name = "all_deps",
-        deps = [
+        exported_deps = [
             "//executorch/kernels/portable/cpu/pattern:pattern",
             "//executorch/kernels/portable/cpu/pattern:bitwise_op",
             "//executorch/kernels/portable/cpu/pattern:comparison_op",
@@ -58,7 +58,7 @@ def define_common_targets():
             "pattern.h",
         ],
         compiler_flags = ["-Wno-missing-prototypes"],
-        deps = [
+        exported_deps = [
             "//executorch/kernels/portable/cpu/util:broadcast_util",
             "//executorch/kernels/portable/cpu/util:functional_util",
             "//executorch/runtime/kernel:kernel_includes",

kernels/portable/cpu/selective_build.h

@@ -13,7 +13,7 @@
 #ifdef EXECUTORCH_SELECTIVE_BUILD_DTYPE
 // include header generated by
 // executorch/codegen/tools/gen_selected_op_variants.py
-#include <executorch/kernels/portable/cpu/selected_op_variants.h>
+#include "selected_op_variants.h"
 #else
 // dummy implementation
 inline constexpr bool should_include_kernel_dtype(

kernels/portable/cpu/targets.bzl

@@ -46,7 +46,11 @@ def define_common_targets():
         ],
         srcs = [],
         exported_headers = ["vec_ops.h"],
-        visibility = ["//executorch/kernels/portable/cpu/...", "//executorch/kernels/quantized/..."],
+        visibility = [
+            "//executorch/kernels/portable/...",
+            "//executorch/kernels/quantized/...",
+            "@EXECUTORCH_CLIENTS",
+        ],
     )
 
     # Only for use by targets in this directory. Defines constants like M_PI
@@ -58,7 +62,7 @@ def define_common_targets():
             "math_constants.h",
         ],
         visibility = [
-            "//executorch/kernels/portable/cpu/...",
+            "//executorch/kernels/portable/...", "@EXECUTORCH_CLIENTS",
         ],
     )
 

kernels/portable/cpu/util/targets.bzl

@@ -11,7 +11,7 @@ def define_common_targets():
     # build, where the portable ops are built from source and linked with :all_deps
     runtime.cxx_library(
         name = "all_deps",
-        deps = [
+        exported_deps = [
             "//executorch/extension/threadpool:threadpool",
             "//executorch/kernels/portable/cpu/util:functional_util",
             "//executorch/kernels/portable/cpu/util:broadcast_util",

shim_et/xplat/executorch/codegen/codegen.bzl

@@ -1,7 +1,7 @@
 load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "get_default_executorch_platforms", "is_xplat", "runtime", "struct_to_json")
 load("@fbsource//xplat/executorch/build:selects.bzl", "selects")
-load("@fbsource//xplat/executorch/kernels/portable:op_registration_util.bzl", "portable_header_list", "portable_source_list")
-load("@fbsource//xplat/executorch/kernels/optimized:op_registration_util.bzl", "optimized_header_list", "optimized_source_list")
+load("@fbsource//xplat/executorch/kernels/portable:op_registration_util.bzl", "portable_source_list")
+load("@fbsource//xplat/executorch/kernels/optimized:op_registration_util.bzl", "optimized_source_list")
 load(
     "@fbsource//xplat/executorch/kernels/optimized:lib_defs.bzl",
     "get_vec_deps",
@@ -410,26 +410,15 @@ def copy_files(genrule_name, target, file_list):
 def build_portable_header_lib(name, oplist_header_name, feature = None):
     """Build the portable headers into a header-only library.
     Ensures that includes work across portable and optimized libs.
-    #include "executorch/kernels/portable/cpu/<header.h>"
     """
-    # Copy portable header files.
-    portable_header_files = {}
-    genrule_name = name + "_copy_portable_header"
-    copy_files(genrule_name, "//executorch/kernels/portable/cpu:portable_header_files", portable_header_list())
-    for header in portable_header_list():
-        portable_header_files[header] = ":{}[{}]".format(genrule_name, header)
-
-    # Include dtype header.
-    portable_header_files["selected_op_variants.h"] = ":{}[selected_op_variants]".format(oplist_header_name)
-
-    # Build portable headers lib.
     runtime.cxx_library(
         name = name,
         srcs = [],
-        exported_headers = portable_header_files,
+        exported_headers = {
+            "selected_op_variants.h":":{}[selected_op_variants]".format(oplist_header_name),
+        },
         exported_preprocessor_flags = ["-DEXECUTORCH_SELECTIVE_BUILD_DTYPE"],
-        # header_namespace is only available in xplat. See https://fburl.com/code/we2gvopk
-        header_namespace = "executorch/kernels/portable/cpu",
+        header_namespace = "",
         feature = feature,
     )
 
@@ -464,9 +453,13 @@ def build_portable_lib(name, oplist_header_name, portable_header_lib, feature =
         name = name,
         srcs = portable_source_files,
         exported_preprocessor_flags = ["-DEXECUTORCH_SELECTIVE_BUILD_DTYPE"],
-        deps = ["//executorch/kernels/portable/cpu/pattern:all_deps", "//executorch/kernels/portable/cpu/util:all_deps"] +  [":" + portable_header_lib],
-        # header_namespace is only available in xplat. See https://fburl.com/code/we2gvopk
-        header_namespace = "executorch/kernels/portable/cpu",
+        deps = [
+            "//executorch/kernels/portable/cpu:math_constants",
+            "//executorch/kernels/portable/cpu:scalar_utils",
+            "//executorch/kernels/portable/cpu:vec_ops",
+            "//executorch/kernels/portable/cpu/pattern:all_deps",
+            "//executorch/kernels/portable/cpu/util:all_deps",
+        ] +  [":" + portable_header_lib],
         compiler_flags = compiler_flags,
         # WARNING: using a deprecated API to avoid being built into a shared
         # library. In the case of dynamically loading so library we don't want
@@ -492,13 +485,6 @@ def build_optimized_lib(name, oplist_header_name, portable_header_lib, feature =
     for op in optimized_source_list():
         optimized_source_files.append(":{}[{}]".format(source_genrule, op))
 
-    # Copy optimized header files.
-    optimized_header_files = {}
-    header_genrule = name + "_copy_optimized_header"
-    copy_files(header_genrule, "//executorch/kernels/optimized/cpu:optimized_header_files", optimized_header_list())
-    for header in optimized_header_list():
-        optimized_header_files[header] = ":{}[{}]".format(header_genrule, header)
-
     # For shared library build, we don't want to expose symbols of
     # kernel implementation (ex torch::executor::native::tanh_out)
     # to library users. They should use kernels through registry only.
@@ -514,6 +500,13 @@ def build_optimized_lib(name, oplist_header_name, portable_header_lib, feature =
         compiler_flags += ["-fvisibility=hidden"]
 
     # Set up dependencies.
+    portable_lib_deps = [
+        "//executorch/runtime/core/portable_type/c10/c10:aten_headers_for_executorch",
+        "//executorch/runtime/kernel:kernel_includes",
+        "//executorch/kernels/portable/cpu:scalar_utils",
+        "//executorch/kernels/portable/cpu/pattern:all_deps",
+        "//executorch/kernels/portable/cpu/util:all_deps",
+    ]
     optimized_lib_deps = [
         "//executorch/kernels/optimized/cpu:add_sub_impl",
         "//executorch/kernels/optimized/cpu:binary_ops",
@@ -522,21 +515,14 @@ def build_optimized_lib(name, oplist_header_name, portable_header_lib, feature =
         "//executorch/kernels/optimized:libblas",
         "//executorch/kernels/optimized:libutils",
         "//executorch/kernels/optimized:libvec",
-        "//executorch/kernels/portable/cpu/pattern:all_deps", 
-        "//executorch/kernels/portable/cpu/util:all_deps",
-        "//executorch/runtime/core/portable_type/c10/c10:aten_headers_for_executorch",
-        "//executorch/runtime/kernel:kernel_includes",
-        ":" + portable_header_lib,
     ] + get_vec_deps()
-    
+
     # Build optimized lib.
     runtime.cxx_library(
         name = name,
         srcs = optimized_source_files,
         exported_preprocessor_flags = ["-DEXECUTORCH_SELECTIVE_BUILD_DTYPE"],
-        deps = optimized_lib_deps,
-        # header_namespace is only available in xplat. See https://fburl.com/code/we2gvopk
-        header_namespace = "executorch/kernels/optimized/cpu",
+        deps = portable_lib_deps + optimized_lib_deps + [":" + portable_header_lib],
         compiler_flags = compiler_flags,
         preprocessor_flags = get_vec_preprocessor_flags(),
         # sleef needs to be added as a direct dependency of the operator target when building for Android,
@@ -627,21 +613,46 @@ def executorch_generated_lib(
         deps: Additinal deps of the main C++ library. Needs to be in either `//executorch` or `//caffe2` module.
         platforms: platforms args to runtime.cxx_library (only used when in xplat)
         manual_registration: if true, generate RegisterKernels.cpp and RegisterKernels.h.
-        use_default_aten_ops_lib: If `aten_mode` is True AND this flag is True, use `torch_mobile_all_ops_et` for ATen operator library.
+        use_default_aten_ops_lib: If `aten_mode` is True AND this flag is True,
+            use `torch_mobile_all_ops_et` for ATen operator library.
         xplat_deps: Additional xplat deps, can be used to provide custom operator library.
         fbcode_deps: Additional fbcode deps, can be used to provide custom operator library.
         compiler_flags: compiler_flags args to runtime.cxx_library
-        dtype_selective_build: In additional to operator selection, dtype selective build further selects the dtypes for each operator. Can be used with model or dict selective build APIs, where dtypes can be specified. Note: this is only available in xplat.
-        feature: Product-Feature Hierarchy (PFH). For internal use only, required for FoA in production. See: https://fburl.com/wiki/2wzjpyqy
-        support_exceptions: enable try/catch wrapper around operator implemntations to make sure exceptions thrown will not bring down the process. Disable if your use case disables exceptions in the build.
+        dtype_selective_build: In additional to operator selection, dtype selective build
+            further selects the dtypes for each operator. Can be used with model or dict
+            selective build APIs, where dtypes can be specified.
+        feature: Product-Feature Hierarchy (PFH). For internal use only, required
+            for FoA in production. See: https://fburl.com/wiki/2wzjpyqy
+        expose_operator_symbols: By default, fvisibility=hidden is set for executorch kernel
+            libraries built with dtype selective build. This options removes the compiler
+            flag and allows operators to be called outside of the kernel registry.
+            NOTE: It is not recommended to set this to True, as symbols may clash (duplicate
+            symbols errors) if multiple executorch_generated_libs are included by a parent library.
+        support_exceptions: enable try/catch wrapper around operator implementations
+            to make sure exceptions thrown will not bring down the process. Disable if your
+            use case disables exceptions in the build.
     """
     if functions_yaml_target and aten_mode:
         fail("{} is providing functions_yaml_target in ATen mode, it will be ignored. `native_functions.yaml` will be the source of truth.".format(name))
 
     if not aten_mode and not functions_yaml_target and not custom_ops_yaml_target:
         fail("At least one of functions_yaml_target, custom_ops_yaml_target needs to be provided")
 
+    if expose_operator_symbols:
+        if not dtype_selective_build:
+            fail("""
+            expose_operator_symbols is only available in dtype selective build mode.
+            See: https://www.internalfb.com/wiki/PyTorch/Teams/Edge/PyTorch_Edge_Core_Team/Dtype_Selective_Build/""")
+
     if dtype_selective_build:
+        if not expose_operator_symbols and not is_xplat():
+            fail("""
+                Dtype selective build with expose_operator_symbols=False works only in xplat - 
+                there are undefined symbols otherwise. Please try to use xplat, or talk to the
+                executorch team. Setting expose_operator_symbols=True is not recommended as the
+                exposed symbols may clash (duplicate symbols errors) if multiple
+                excutorch_generated_libs are included by a parent library.""")
+
         if (not "//executorch/kernels/portable:operators" in kernel_deps) and (not "//executorch/kernels/optimized:optimized_operators" in kernel_deps):
             fail("""
             !!WARNING!! Dtype selective build is available for the portable and optimized kernel libraries.
@@ -655,7 +666,7 @@ def executorch_generated_lib(
             If you have a custom kernel library, please remove `dtype_selective_build=True`
             and use regular selective build.
             """.format(kernel_deps))
-        
+
         # Dtype selective build requires that the portable/optimized kernel libraries are not passed into `deps`.
         if ("//executorch/kernels/portable:operators" in kernel_deps):
             index = 0
@@ -757,11 +768,11 @@ def executorch_generated_lib(
 
     portable_lib = []
     optimized_lib = []
-    if dtype_selective_build and is_xplat():
+    if dtype_selective_build:
         # Build portable headers lib. Used for portable and optimized kernel libraries.
         portable_header_lib = name + "_portable_header_lib"
         build_portable_header_lib(portable_header_lib, oplist_header_name, feature)
-        
+
         if "//executorch/kernels/portable:operators" in kernel_deps:
             # Remove portable from kernel_deps as we're building it from source.
             kernel_deps.remove("//executorch/kernels/portable:operators")
@@ -770,11 +781,11 @@ def executorch_generated_lib(
             portable_lib_name = name + "_portable_lib"
             build_portable_lib(portable_lib_name, oplist_header_name, portable_header_lib, feature, expose_operator_symbols)
             portable_lib = [":{}".format(portable_lib_name)]
-            
+
         if "//executorch/kernels/optimized:optimized_operators" in kernel_deps:
             # Remove optimized from kernel_deps as we're building it from source.
             kernel_deps.remove("//executorch/kernels/optimized:optimized_operators")
-            
+
             # Build optimized lib.
             optimized_lib_name = name + "_optimized_lib"
             build_optimized_lib(optimized_lib_name, oplist_header_name, portable_header_lib, feature, expose_operator_symbols)

shim_et/xplat/executorch/kernels/optimized/op_registration_util.bzl

@@ -269,7 +269,3 @@ OPTIMIZED_ATEN_OPS = (
 def optimized_source_list():
     """All the source file names from //executorch/kernels/optimized/cpu"""
     return [op["name"] + ".cpp" for op in OPTIMIZED_ATEN_OPS]
-
-def optimized_header_list():
-    """All the header file names from //executorch/kernels/optimized/cpu"""
-    return ["binary_ops.h", "fft_utils.h", "moments_utils.h", "op_add_sub_impl.h",]

shim_et/xplat/executorch/kernels/portable/op_registration_util.bzl

@@ -1333,7 +1333,3 @@ CUSTOM_OPS = (
 def portable_source_list():
     """All the source file names from //executorch/kernels/portable/cpu/"""
     return [op["name"] + ".cpp" for op in ATEN_OPS + CUSTOM_OPS]
-
-def portable_header_list():
-    """All the header file names from //executorch/kernels/portable/cpu/"""
-    return ["selective_build.h", "scalar_utils.h", "math_constants.h", "vec_ops.h"]