simplify tensors before scheduling [pr] (#8580)

* delete forced_realize

* put that back

* work

* remove forced_realize

* expectedFailures

* contiguous(buffer)

* multi

* expectedFailures

* cleaner create_subbuffer

* more comments

* remove that

* note

* realizes

* work

* one upat and image is back

* remove

* cleaner

* fix test_complex_backward for now

---------

Co-authored-by: George Hotz <geohot@gmail.com>

test/test_image_dtype.py

@@ -113,7 +113,8 @@ class TestImageDType(unittest.TestCase):
     assert it.lazydata.base.realized._buf != b1
 
   # issue caused by: don't realize image to image casts. this is part of a larger problem
-  @unittest.expectedFailure
+  #@unittest.expectedFailure
+  # update: passing after tensor_map
   def test_lil_model(self):
     with Context(IMAGE=2):
       x = Tensor.zeros(1, 1)

test/test_schedule.py

@@ -220,7 +220,7 @@ class TestSchedule(unittest.TestCase):
     GlobalCounters.reset()
     expr = (a*b)/b
     expr.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    self.assertEqual(GlobalCounters.kernel_count, 0) # the scheduler can fold divs now!
     self.assertEqual(GlobalCounters.global_ops, 0)
     np.testing.assert_allclose(expr.numpy(), np.full((4,), 4.0))
 
@@ -229,7 +229,7 @@ class TestSchedule(unittest.TestCase):
     GlobalCounters.reset()
     expr = a/a
     expr.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
     self.assertEqual(GlobalCounters.global_ops, 0)
     np.testing.assert_allclose(expr.numpy(), np.full((4,), 1.0))
 
@@ -2204,7 +2204,7 @@ class TestConst(unittest.TestCase):
     sched = add.schedule()
     self.assertEqual(len(sched), 0)
     # b+0 and b share the same underlying device memory
-    self.assertIs(add.lazydata.realized, b.lazydata.realized)
+    self.assertIs(add.lazydata.buffer, b.lazydata.buffer)
     self.assertListEqual(add.tolist(), [2, 2, 2, 2])
 
   def test_src_masked_const_folding(self):