Update251203 (#233)

tinygrad_repo/extra/optimization/generate_dataset.sh

@@ -7,7 +7,7 @@ rm $LOGOPS
 test/external/process_replay/reset.py
 
 CI=1 python3 -m pytest -n=auto test/test_ops.py test/test_nn.py test/test_winograd.py test/models/test_real_world.py --durations=20
-GPU=1 python3 -m pytest test/test_tiny.py
+CL=1 python3 -m pytest test/test_tiny.py
 
 # extract, sort and uniq
 extra/optimization/extract_dataset.py

tinygrad_repo/extra/optimization/helpers.py

@@ -1,6 +1,6 @@
 # stuff needed to unpack a kernel
 from tinygrad import Variable
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.dtype import dtypes, PtrDType
 from tinygrad.shape.shapetracker import ShapeTracker
@@ -115,7 +115,7 @@ def time_linearizer(lin:Kernel, rawbufs:list[Buffer], allow_test_size=True, max_
   assert dev.compiler is not None
 
   rawbufs = _ensure_buffer_alloc(rawbufs)
-  var_vals: dict[Variable, int] = {k:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
+  var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
   p = get_program(lin.get_optimized_ast(), lin.opts)
   tms = _time_program(p, dev.compiler.compile(p.src), var_vals, rawbufs,
                       max_global_size=max_global_size if allow_test_size else None, clear_l2=clear_l2, cnt=cnt, name=to_function_name(lin.name))

tinygrad_repo/extra/optimization/test_beam_search.py

@@ -16,9 +16,9 @@ class TestBeamSearch(unittest.TestCase):
     BEAM.value = self.old_beam
 
   def test_variable_ast_beam(self):
-    with Context(IGNORE_OOB=1):
-      a = rand(3, 3).reshape((Variable("a", 1, 10).bind(3), 3))
-      a = (a+1).realize()
+    vi = Variable("a", 1, 10).bind(3)
+    a = rand(10, 3)[:vi]
+    a = (a+1).realize()
 
   def test_big_prime_number(self):
     a = rand(367, 367)
@@ -42,18 +42,16 @@ class TestBeamSearch(unittest.TestCase):
 
   def test_variable_big_prime_number(self):
     v = Variable("v", 1, 400).bind(367)
-    a = rand(367, 367)
-    b = rand(367, 367)
-    with Context(IGNORE_OOB=1):
-      c = (a.reshape(367, v) @ b.reshape(v, 367)).realize()
-      np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), atol=1e-4, rtol=1e-4)
+    a = rand(367, 400)
+    b = rand(400, 367)
+    c = (a[:, :v] @ b[:v, :]).realize()
+    np.testing.assert_allclose(c.numpy(), a[:, :367].numpy() @ b[:367, :].numpy(), atol=1e-4, rtol=1e-4)
 
   def test_variable_shrink_prime_number(self):
     v = Variable("v", 1, 400).bind(367)
     a = rand(400, 367)
-    with Context(IGNORE_OOB=1):
-      b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
-      np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)
+    b = (a.shrink(((0,v), None))+1)[:367,:367].realize()
+    np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)
 
   def test_no_mutate_rawbuffers(self):
     a = rand(3, 3).realize()