matmul: change onednn commit & revert wa (#3556)

Author: toyxu

csrc/gpu/oneDNN/Matmul.h

@@ -23,26 +23,10 @@ namespace oneDNN {
 static inline void matmul(
     Tensor& result,
     const Tensor& mat1,
-    const Tensor& mat2_,
+    const Tensor& mat2,
     const Tensor& b_raw,
     bool m2_trans,
     Attr attr) {
-  // FIXME: There is a onednn bug causes broadcast post fusion to be
-  // accuracy error when t2 is column major
-  Tensor mat2;
-  DeviceId curDevID;
-  dpcppGetDevice(&curDevID);
-  bool mat2_wa_flag =
-      (!Settings::I().has_2d_block_array(curDevID)) && (mat2_.dim() == 2);
-  if (mat2_wa_flag && m2_trans == false && mat2_.size(0) < 16) {
-    mat2 = mat2_.transpose(0, 1).contiguous();
-    m2_trans = true;
-  } else if (mat2_wa_flag && m2_trans == true && mat2_.size(1) < 16) {
-    mat2 = mat2_.contiguous();
-  } else {
-    mat2 = mat2_;
-  }
-
   size_t dims = result.dim();
   TORCH_CHECK(
       dims == 2 || dims == 3,