Remove fbgemm autovec (#4123)

Summary:

DO NOT LAND

Testing PyTorch CI without fbgemm autovec

Differential Revision: D74747522

Author: gchalump

CMakeLists.txt

@@ -105,6 +105,7 @@ option(FBGEMM_BUILD_TESTS "Build fbgemm unit tests" ON)
 option(FBGEMM_BUILD_BENCHMARKS "Build fbgemm benchmarks" ON)
 option(FBGEMM_BUILD_DOCS "Build fbgemm documentation" OFF)
 option(FBGEMM_BUILD_FBGEMM_GPU "Build fbgemm_gpu library" OFF)
+option(DISABLE_FBGEMM_AUTOVEC "Disable FBGEMM Autovec" OFF)
 
 if(FBGEMM_BUILD_TESTS)
   enable_testing()
@@ -422,3 +423,11 @@ endif()
 if(FBGEMM_BUILD_FBGEMM_GPU)
   add_subdirectory(fbgemm_gpu)
 endif()
+
+if(HAVE_GNU_F2H_IEEE)
+    add_definitions(-DHAVE_GNU_F2H_IEEE)
+endif()
+
+if(DISABLE_FBGEMM_AUTOVEC)
+  add_definitions(-DDISABLE_FBGEMM_AUTOVEC)
+endif()

cmake/modules/FindGnuH2fIeee.cmake

@@ -0,0 +1,21 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+################################################################################
+# Finds and sets GNU_FH2_IEEE compilation flags
+################################################################################
+
+INCLUDE(CheckCXXSourceCompiles)
+
+CHECK_CXX_SOURCE_COMPILES("
+    #include <arm_neon.h>
+    int main() {
+        float f = 1.0f;
+        uint16_t h = __gnu_f2h_ieee(f);
+        return 0;
+    }
+" HAVE_GNU_F2H_IEEE)

fbgemm_gpu/cmake/Fbgemm.cmake

@@ -10,14 +10,17 @@
 
 set(fbgemm_sources_normal
   "${FBGEMM}/src/EmbeddingSpMDM.cc"
-  "${FBGEMM}/src/EmbeddingSpMDMAutovec.cc"
   "${FBGEMM}/src/EmbeddingSpMDMNBit.cc"
   "${FBGEMM}/src/QuantUtils.cc"
   "${FBGEMM}/src/RefImplementations.cc"
   "${FBGEMM}/src/RowWiseSparseAdagradFused.cc"
   "${FBGEMM}/src/SparseAdagrad.cc"
   "${FBGEMM}/src/Utils.cc")
 
+if(NOT DISABLE_FBGEMM_AUTOVEC)
+  list(APPEND fbgemm_sources_normal "${FBGEMM}/src/EmbeddingSpMDMAutovec.cc")
+endif()
+
 set(fbgemm_sources_avx2
   "${FBGEMM}/src/EmbeddingSpMDMAvx2.cc"
   "${FBGEMM}/src/QuantUtilsAvx2.cc")

include/fbgemm/FloatConversion.h

@@ -287,7 +287,7 @@ inline float cpu_half2float_ref(const float16 h) {
 // Same as the previous function, but use the built-in fp16 to fp32
 // conversion provided by the compiler
 inline float cpu_half2float(const float16 h) {
-#ifdef HAS_NATIVE_FP16_TYPE
+#if defined(HAS_NATIVE_FP16_TYPE) && defined(HAVE_GNU_F2H_IEEE)
   __fp16 h_fp16;
   std::memcpy(&h_fp16, &h, sizeof(__fp16));
   return h_fp16;
@@ -297,7 +297,7 @@ inline float cpu_half2float(const float16 h) {
 }
 
 inline float16 cpu_float2half(const float f) {
-#ifdef HAS_NATIVE_FP16_TYPE
+#if defined(HAS_NATIVE_FP16_TYPE) && defined(HAVE_GNU_F2H_IEEE)
   __fp16 h = f;
   float16 res;
   std::memcpy(&res, &h, sizeof(__fp16));

src/EmbeddingSpMDMAutovec.h

@@ -14,7 +14,9 @@
 
 #include "fbgemm/FbgemmEmbedding.h"
 
+#ifndef DISABLE_FBGEMM_AUTOVEC
 #define FBGEMM_AUTOVEC_AVAILABLE
+#endif
 
 namespace fbgemm {