Remove AVX compilation on aarch64 builds

Summary:
X-link: facebookresearch/FBGEMM#2056

We are no longer using SIMDE to build AVX code targeting aarch64

Build times should decrease

Reviewed By: mcfi

Differential Revision: D85361586

Author: Nicoshev

src/TransposeUtils.cc

@@ -57,11 +57,15 @@ void transpose_simd(
 #else
   static const auto iset = fbgemmInstructionSet();
   // Run time CPU detection
-#if defined(FBGEMM_FBCODE) || !defined(__aarch64__)
+#if defined(FBGEMM_FBCODE) && (defined(__AVX512F__) || defined(__AVX2__))
   if (isZmm(iset)) {
+#ifdef __AVX512F__
     internal::transpose_avx512<T>(M, N, src, ld_src, dst, ld_dst);
+#endif
   } else if (isYmm(iset)) {
+#ifdef __AVX2__
     internal::transpose_avx2<T>(M, N, src, ld_src, dst, ld_dst);
+#endif
   } else
 #endif
   {

src/UtilsAvx2.cc

@@ -15,6 +15,8 @@
 
 namespace fbgemm::internal {
 
+#ifdef __AVX2__
+
 template <>
 void transpose_avx2(
     int64_t M,
@@ -336,4 +338,6 @@ void transpose_avx2(
   }
 }
 
+#endif // __AVX2__
+
 } // namespace fbgemm::internal

src/UtilsAvx512.cc

@@ -16,6 +16,8 @@ namespace fbgemm {
 
 namespace {
 
+#ifdef __AVX512F__
+
 // 16 * 6 = 96 instructions
 inline void transpose_kernel_16x16_avx512(
     const float* src,
@@ -2441,6 +2443,8 @@ void transpose_avx512(
   }
 }
 
+#endif // __AVX512F__
+
 } // namespace internal
 
 } // namespace fbgemm