Update on "Mostly sync BlasKernel.cpp with ATen ReducedPrecisionGemvFastPathKernel"

The two files were similar, but diverged due to recent changes. Since we have sharing of PyTorch headers, we can keep them mostly the same; differences are some of the namespace stuff and a couple of EXECUTORCH NOTEs.

Differential Revision: [D74702689](https://our.internmc.facebook.com/intern/diff/D74702689/)

[ghstack-poisoned]

Author: swolchok

kernels/optimized/blas/BlasKernel.cpp

@@ -95,10 +95,13 @@ float reduce(vec::VectorizedN<float, kF32RegistersPerIteration>& x) {
   return reduce(x[0]);
 }
 
+// EXECUTORCH NOTE: removed __ARM_FEATURE_BF16_VECTOR_ARITHMETIC gate
+// added in https://github.com/pytorch/pytorch/pull/152766, which I
+// complained on.
+
 // We would have to write a separate SVE-specific path to use SVE
 // BFDOT. Deferring that for now to get the NEON/ASIMD BFDOT path
 // working.
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
 #if defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE) && defined(__clang__) && __clang_major__ > 15
 // https://godbolt.org/z/z8P4Yncra
 #define COMPILER_SUPPORTS_BF16_TARGET 1
@@ -109,9 +112,6 @@ float reduce(vec::VectorizedN<float, kF32RegistersPerIteration>& x) {
 #else // defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE) && defined(__clang__) && __clang_major__ > 15
 #define COMPILER_SUPPORTS_BF16_TARGET 0
 #endif // defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE) && defined(__clang__) && __clang_major__ > 15
-#else // __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-#define COMPILER_SUPPORTS_BF16_TARGET 0
-#endif // __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
 
 #if COMPILER_SUPPORTS_BF16_TARGET
 #define TARGET_ARM_BF16_ATTRIBUTE __attribute__((target("arch=armv8.2-a+bf16")))

kernels/optimized/lib_defs.bzl

@@ -2,10 +2,6 @@ load("@fbsource//tools/build_defs:default_platform_defs.bzl", "DEVSERVER_PLATFOR
 load("@fbsource//tools/build_defs:fb_native_wrapper.bzl", "fb_native")
 load("@fbsource//xplat/executorch/backends/xnnpack/third-party:third_party_libs.bzl", "third_party_dep")
 load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
-load(
-    "@fbsource//xplat/executorch/kernels/portable:op_registration_util.bzl",
-    "get_compiler_optimization_flags",
-)
 
 # Because vec exists as a collection of header files, compile and preprocessor
 # flags applied to the vec target do not have any effect, since no compilation
@@ -200,7 +196,12 @@ def define_libs(is_fbcode=False):
             exported_headers = native.glob([
                 "blas/**/*.h",
             ]),
-            compiler_flags = ["-Wno-pass-failed"] + get_compiler_optimization_flags(),
+            compiler_flags = ["-Wno-pass-failed"] + select({
+                "ovr_config//runtime:fbcode": [],
+                # TODO: replace with get_compiler_optimization_flags from op_registration_util.bzl when that
+                # is re-enabled.
+                "DEFAULT": ["-Os"],
+            }),
             header_namespace = "executorch/kernels/optimized",
             visibility = [
                 "//executorch/...",