Fast transposed copy

Author: adamlerer

test/test_torch.py

@@ -3230,6 +3230,14 @@ def test_Size(self):
         self.assertIsInstance(x[:-1], torch.Size)
         self.assertIsInstance(x + x, torch.Size)
 
+    # unit test for THTensor_(copyTranspose)
+    @unittest.skipIf(not TEST_NUMPY, "Numpy not found")
+    def test_big_transpose(self):
+        t = torch.rand(456, 789)
+        t1 = t.t().contiguous()
+        t2 = torch.from_numpy(t.numpy().transpose())
+        self.assertEqual(t1, t2)
+
 # Functions to test negative dimension wrapping
 METHOD = 1
 INPLACE_METHOD = 2

torch/csrc/generic/Tensor.cpp

@@ -816,6 +816,8 @@ PyTypeObject THPTensorStatelessType = {
 void THPTensor_(initCopyMethods)()
 {
   auto& h = THTensor_(copy_functions);
+  // copy from same type
+  THPInsertCopyFunction(h, &THTensor_(copy));
   // copy from CPU types
   THPInsertCopyFunction(h, &THTensor_(copyByte));
   THPInsertCopyFunction(h, &THTensor_(copyChar));