Update on "inductor: make onednn linear inputs are always real contig…

…uous"



For OneDNN linear, if packed linear inputs are not the default contiguous tensor, it always calls in ref pat and gets a worse performance, this PR will force its inputs to the actual default contiguous tensor.

cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 ipiszy ngimel yf225 chenyang78 kadeng muchulee8 aakhundov

[ghstack-poisoned]

test/inductor/test_cpu_repro.py

@@ -2472,7 +2472,7 @@ def fn(x, y):
     def test_linear_with_no_default_contiguous_input(self):
         mod = torch.nn.Sequential(torch.nn.Linear(16, 16)).eval()
         temp = torch.randn(1, 16, 1, 1)
-        v = torch.ops.inductor._reinterpret_tensor(temp, [1, 16], [0, 1], 0)
+        v = torch.as_strided(temp, [1, 16], [0, 1], 0)
         self.assertTrue(v.is_contiguous())
         with torch.no_grad():
             self.common(

torch/_inductor/fx_passes/mkldnn_fusion.py

@@ -1006,7 +1006,6 @@ def linear(match, *args, **kwargs):
                     if free_symbols(batch_size)
                     else batch_size,
                 )
-                packed_weight_inputs = (transpose_weight_node, batch_size)
                 packed_weight_op = (
                     mkldnn._reorder_linear_weight
                     if is_bf16_weight