Use std::fma in tail of Fused8BitRowwiseQuantizedSBFloatToFloatOrHalf…

…Avx2 (#1191) Summary: Pull Request resolved: #1191 Another missing FMA. Reviewed By: jspark1105 Differential Revision: D37670913 fbshipit-source-id: da0a95d4e1dff8d6015a2d5fd398f56e945bd17a
pytorch · Jul 8, 2022 · 56cd603 · 56cd603
1 parent 64a5c4a
commit 56cd603
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/src/QuantUtilsAvx2.cc b/src/QuantUtilsAvx2.cc
@@ -2164,8 +2164,8 @@ void Fused8BitRowwiseQuantizedSBFloatToFloatOrHalfAvx2(
     }
 
     for (; col < output_columns; ++col) {
-      float output_value =
-          input_row[col] * input_row_scale_bias[0] + input_row_scale_bias[1];
+      float output_value = std::fma(
+          input_row[col], input_row_scale_bias[0], input_row_scale_bias[1]);
       if (std::is_same<OutputType, float>()) {
         output_row[col] = output_value;
       } else {