[fbgemm_gpu[ Fix aarch64 build issues caused by D99968947 (pytorch#5655)

Summary:
Pull Request resolved: pytorch#5655

- Fix aarch64 build issues caused by D99968947

Reviewed By: Nicoshev, ZhijianJiang

Differential Revision: D101387235

fbshipit-source-id: 1cb3c11f4224571e0706226a7617a58ac304d8ac

Author: q10

include/fbgemm/FbgemmConvert.h

@@ -118,11 +118,11 @@ FBGEMM_API void FloatToFloat16_simd(
 FBGEMM_API void
 Float16ToFloat_simd(const float16* src, float* dst, size_t size);
 
+#if !defined(__aarch64__)
 /**
  * @brief AVX2 implementation to convert fp32 numbers to fp16 numbers.
  *
  */
-#if !defined(__aarch64__)
 FBGEMM_API void FloatToFloat16_avx2(
     const float* src,
     float16* dst,

src/QuantUtilsAvx2.cc

@@ -1626,7 +1626,11 @@ void FloatOrHalfToFusedNBitRowwiseQuantizedSBHalfAvx2(
       maximum_element = min_max_row_float[1];
 
       if constexpr (std::is_same_v<InputType, float16>) {
+#if !defined(__aarch64__)
         Float16ToFloat_avx2(input_row, input_row_float_for_fp16, input_columns);
+#else
+        Float16ToFloat_ref(input_row, input_row_float_for_fp16, input_columns);
+#endif
       }
     } else {
       __m256 min_v = _mm256_set1_ps(minimum_element);
@@ -1847,7 +1851,11 @@ void FloatOrHalfToFused8BitRowwiseQuantizedSBFloatAvx2(
       maximum_element = min_max_row_float[1];
 
       if constexpr (std::is_same_v<InputType, float16>) {
+#if !defined(__aarch64__)
         Float16ToFloat_avx2(input_row, input_row_float_for_fp16, input_columns);
+#else
+        Float16ToFloat_ref(input_row, input_row_float_for_fp16, input_columns);
+#endif
       }
     } else {
       __m256 min_v = _mm256_set1_ps(minimum_element);