do not use getenv('PTX') in tests (#12095)

* test without ptx

* fix tests

* fix test

* linters

test/test_arange.py

@@ -6,6 +6,7 @@ from tinygrad.engine.realize import run_schedule
 from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
 from tinygrad.uop.ops import Ops
+from tinygrad.renderer.ptx import PTXRenderer
 
 class TestArange(unittest.TestCase):
   def _get_flops(self, N, opts=None):
@@ -26,7 +27,7 @@ class TestArange(unittest.TestCase):
     print(f"{f1=}, {f2=}")
     # add 1 to avoid divide by 0. arange is 0 flops now!
     assert (f1 < 6000 and f2 < 6000) or ((f2+1) / (f1+1) < 16), f"bad complexity, flops {(f2+1) / (f1+1):.1f}X while inputs 10X"
-    if limit is not None and not getenv("PTX"):
+    if limit is not None and not isinstance(Device[Device.DEFAULT].renderer, PTXRenderer):
       # PTX counts index ALU in flops
       assert f1 <= limit, f"{f1=}, {limit=}"
 

test/test_dtype.py

@@ -5,6 +5,7 @@ from typing import Any, List
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv, DEBUG, CI
 from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8, _to_np_dtype, _to_torch_dtype
+from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad import Device, Tensor, dtypes
 from hypothesis import assume, given, settings, strategies as strat
 from test.helpers import rand_for_dtype
@@ -49,7 +50,7 @@ def _test_cast(a:Tensor, target_dtype:DType):
 
   _test_op(lambda: a.cast(target_dtype), target_dtype, list(a.numpy().astype(_to_np_dtype(target_dtype))))
 def _test_bitcast(a:Tensor, target_dtype:DType, target=None):
-  if getenv("PTX") and a.dtype == dtypes.int8 and target_dtype.itemsize != a.dtype.itemsize:
+  if isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and a.dtype == dtypes.int8 and target_dtype.itemsize != a.dtype.itemsize:
     raise unittest.SkipTest("shape changing bitcast of int8 broken on PTX")
   expected = torch.tensor(a.tolist(), dtype=_to_torch_storage_type(a.dtype)).view(_to_torch_dtype(target_dtype))
   _test_op(lambda: a.bitcast(target_dtype), target_dtype, target or expected.tolist())
@@ -100,7 +101,7 @@ class TestDType(unittest.TestCase):
     ))
 
   @unittest.skipIf(Device.DEFAULT == "PYTHON", "skip for now")
-  @unittest.skipIf(getenv("PTX"), "skip for now")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "skip for now")
   def test_uint_overflow(self):
     if not dtypes.is_unsigned(self.DTYPE): raise unittest.SkipTest("only for unsigned")
     v = dtypes.max(self.DTYPE)
@@ -255,7 +256,8 @@ class TestFloatDType(TestDType):
 
 class TestDoubleDType(TestDType):
   DTYPE = dtypes.double
-  @unittest.skipIf((CI and Device.DEFAULT in {"CUDA", "NV"}) or getenv("PTX"), "conversion not supported on CI CUDA and PTX")  # TODO: why not?
+  @unittest.skipIf((CI and Device.DEFAULT in {"CUDA", "NV"}) or \
+                    isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "conversion not supported on CI CUDA and PTX")  # TODO: why not?
   def test_float64_increased_precision(self):
     for func in [
       lambda t: t.exp(),
@@ -279,21 +281,21 @@ class TestDoubleDType(TestDType):
 
 class TestInt8DType(TestDType):
   DTYPE = dtypes.int8
-  @unittest.skipIf(getenv("CUDA",0)==1 or getenv("PTX", 0)==1, "cuda saturation works differently")
+  @unittest.skipIf(getenv("CUDA",0)==1 or isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "cuda saturation works differently")
   def test_int8_to_uint8_negative(self):
     _test_op(lambda: Tensor([-1, -2, -3, -4], dtype=dtypes.int8).cast(dtypes.uint8), dtypes.uint8, [255, 254, 253, 252])
 
   def test_int8_to_uint16_negative(self):
     _test_op(lambda: Tensor([-1, -2, -3, -4], dtype=dtypes.int8).cast(dtypes.uint16), dtypes.uint16, [2**16-1, 2**16-2, 2**16-3, 2**16-4])
 
-  @unittest.skipIf(getenv("PTX"), "broken in ptx")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken in ptx")
   def test_bitcast_alt(self):
     a = Tensor([72, -90, 27, 40, -53, 70, 96, 51], dtype=dtypes.int8).bitcast(dtypes.short)
     self.assertListEqual(a.tolist(), [-22968, 10267, 18123, 13152])
 
 class TestUint8DType(TestDType):
   DTYPE = dtypes.uint8
-  @unittest.skipIf(getenv("CUDA",0)==1 or getenv("PTX", 0)==1, "cuda saturation works differently")
+  @unittest.skipIf(getenv("CUDA",0)==1 or isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "cuda saturation works differently")
   def test_uint8_to_int8_overflow(self):
     _test_op(lambda: Tensor([255, 254, 253, 252], dtype=dtypes.uint8).cast(dtypes.int8), dtypes.int8, [-1, -2, -3, -4])
 
@@ -301,7 +303,7 @@ class TestBitCast(unittest.TestCase):
   @given(strat.sampled_from(dtype_ints + dtype_floats), strat.sampled_from(dtype_ints + dtype_floats))
   def test_shape_change_bitcast(self, dt1, dt2):
     # NOTE: this has to be assume to prevent hypothesis from skipping all samples
-    assume(not (getenv("PTX") and dt1 == dtypes.int8)) # TODO: bitcasting int8 fails in PTX
+    assume(not (isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and dt1 == dtypes.int8)) # TODO: bitcasting int8 fails in PTX
     data = rand_for_dtype(dt1, 32).reshape(2, 2, 8)
     expected = torch.tensor(data.tolist(), dtype=_to_torch_storage_type(dt1)).view(_to_torch_dtype(dt2))
     _test_op(lambda: Tensor(data, dtype=dt1).bitcast(dt2), dt2, expected.tolist())

test/test_dtype_alu.py

@@ -5,6 +5,7 @@ from tinygrad.helpers import CI, getenv
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.device import is_dtype_supported
 from tinygrad.runtime.ops_python import from_storage_scalar
+from tinygrad.renderer.ptx import PTXRenderer
 import numpy as np
 import pytest
 from hypothesis import given, strategies as strat, settings, HealthCheck
@@ -91,7 +92,7 @@ def universal_test_midcast(a, b, c, op1, op2, d1:DType, d2:DType):
   an, bn, cn = np.array([a]).astype(_to_np_dtype(d1)), np.array([b]).astype(_to_np_dtype(d1)), np.array([c]).astype(_to_np_dtype(d2))
   tensor_value = op2[0](op1[0](at, bt).cast(d2), ct).numpy()
   numpy_value = op2[1](op1[1](an, bn).astype(_to_np_dtype(d2)), cn)
-  np.testing.assert_allclose(tensor_value, numpy_value, rtol=1e-6 if getenv("PTX") else 1e-7)
+  np.testing.assert_allclose(tensor_value, numpy_value, rtol=1e-6 if isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) else 1e-7)
 
 class TestDTypeALU(unittest.TestCase):
   @unittest.skipUnless(is_dtype_supported(dtypes.float64), f"no float64 on {Device.DEFAULT}")

test/test_linearizer.py

@@ -10,9 +10,10 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import run_schedule, lower_schedule, CompiledRunner, get_program
-from tinygrad.helpers import Context, getenv, flatten, dedup, TC_SELECT, TC_OPT
+from tinygrad.helpers import Context, flatten, dedup, TC_SELECT, TC_OPT
 from tinygrad.dtype import DType, dtypes, PtrDType, AddrSpace
 from tinygrad.codegen import apply_rewrites, rewrites_for_views
+from tinygrad.renderer.ptx import PTXRenderer
 
 class TestLinearizer(unittest.TestCase):
   def test_arg_dedup(self):
@@ -155,7 +156,7 @@ class TestLinearizer(unittest.TestCase):
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
-  @unittest.skipIf(getenv("PTX"), "broken on ptx for some reason")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken on ptx for some reason")
   def test_upcast_with_locals(self):
     x, y = Tensor.rand(1,128), Tensor.rand(128, 128)
     r = (x@y).relu()
@@ -366,7 +367,7 @@ class TestLinearizer(unittest.TestCase):
     helper(Tensor.arange(255), max_ops=2)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
-  @unittest.skipIf(getenv("PTX"), "broken on ptx for some reason")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken on ptx for some reason")
   def test_grouped_store_phis(self):
     """
     float4 acc0 = float4(0.0,0.0,0.0,0.0);
@@ -420,7 +421,7 @@ class TestLinearizer(unittest.TestCase):
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
-  @unittest.skipIf(getenv("PTX"), "broken on ptx for some reason")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken on ptx for some reason")
   def test_grouped_store_local_only(self):
     x, y = Tensor.rand(1,128), Tensor.rand(128, 128)
     r = (x@y).relu()

test/test_linearizer_dumb.py

@@ -6,10 +6,10 @@ import unittest
 from tinygrad import Device, dtypes
 from tinygrad.device import is_dtype_supported
 from tinygrad.uop.ops import UOp, Ops, AxisType, KernelInfo
-from tinygrad.helpers import getenv
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.codegen.opt.search import Opt, OptOps
 from tinygrad.engine.realize import get_program
+from tinygrad.renderer.ptx import PTXRenderer
 
 class TestLinearizerFailure(unittest.TestCase):
   @unittest.expectedFailure
@@ -93,7 +93,7 @@ class TestLinearizerDumb(unittest.TestCase):
 
   @unittest.expectedFailure
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "need float4")
-  @unittest.skipIf(getenv("PTX"), "this is somehow correct in PTX")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "this is somehow correct in PTX")
   def test_upcasted_stores_out_of_order(self):
     c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9360), arg=0, src=())
     c1 = c0.view(ShapeTracker(views=(View(shape=(4, 5, 13, 1, 1, 1, 1, 1, 4, 3, 3), strides=(2340, 468, 36, 0, 0, 0, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=True),)))

test/test_randomness.py

@@ -9,6 +9,7 @@ from tinygrad.device import is_dtype_supported
 from tinygrad.engine.realize import lower_schedule, CompiledRunner
 from hypothesis import given, settings, strategies as strat
 from test.helpers import not_support_multi_device
+from tinygrad.renderer.ptx import PTXRenderer
 
 settings.register_profile("my_profile", max_examples=200, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
 settings.load_profile("my_profile")
@@ -98,7 +99,7 @@ class TestRandomness(unittest.TestCase):
 
     np.testing.assert_allclose(jr, r)
 
-  @unittest.skipIf(getenv("PTX"), "fails with PTX")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "fails with PTX")
   def test_threefry_doesnt_use_long(self):
     for (_,ei) in lower_schedule(Tensor.rand(20).schedule()):
       if isinstance(ei.prg, CompiledRunner):

test/test_tensor.py

@@ -9,7 +9,7 @@ from extra.gradcheck import numerical_jacobian, jacobian, gradcheck
 from hypothesis import given, settings, strategies as strat
 from tinygrad.device import is_dtype_supported
 from tinygrad.uop.ops import Ops, UOp
-from tinygrad.runtime.support.compiler_cuda import PTX
+from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.codegen import full_rewrite
 from tinygrad.dtype import DType
 
@@ -915,7 +915,7 @@ class TestIdxUpcast(unittest.TestCase):
   def test_regular_sym(self):
     self.do_op_then_assert(dtypes.int, 2048, 2048, UOp.variable("dim3", 1, 64).bind(32))
 
-  @unittest.skipIf(PTX, "PTX always convert Ops.INDEX to int64")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX always convert Ops.INDEX to int64")
   def test_symfold(self):
     # This would cause an overflow, but after sym fold it's within int32
     a = Tensor.arange(65535)

test/test_uops.py

@@ -15,6 +15,7 @@ from tinygrad.codegen import full_rewrite
 from tinygrad.uop.symbolic import sym
 from tinygrad.device import is_dtype_supported
 from tinygrad.codegen.opt import Opt, OptOps
+from tinygrad.renderer.ptx import PTXRenderer
 
 def to_uops_list(u:list[UOp], opts=None, skip_check=False) -> list[UOp]: return full_rewrite(UOp.sink(*u), opts)
 
@@ -130,9 +131,9 @@ class TestFloatUOps(TestUOps):
 class TestNonFloatUOps(TestUOps):
   def test_add_int32(self): self._test_bop_fxn(Ops.ADD, lambda a,b: int(a)+int(b), (dtypes.int32, dtypes.int32))
   def test_mul_int32(self): self._test_bop_fxn(Ops.MUL, lambda a,b: int(a)*int(b), (dtypes.int32, dtypes.int32))
-  @unittest.skipUnless(getenv("PTX"), "only ptx uses bitshifts")
+  @unittest.skipUnless(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "only ptx uses bitshifts")
   def test_shr_int32(self): self._test_bop_fxn(Ops.SHR, lambda a,b: int(a)>>int(b), (dtypes.int32, dtypes.int32), no_b_neg=True)
-  @unittest.skipUnless(getenv("PTX"), "only ptx uses bitshifts")
+  @unittest.skipUnless(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "only ptx uses bitshifts")
   def test_shl_int32(self): self._test_bop_fxn(Ops.SHL, lambda a,b: int(a)<<int(b), (dtypes.int32, dtypes.int32), no_b_neg=True)
   def test_div_int32(self):
     self._test_bop_fxn(Ops.IDIV, lambda a,b: int(a/b), (dtypes.int32, dtypes.int32), no_b_zero=True)
@@ -370,7 +371,7 @@ class TestLocalAccess(unittest.TestCase):
     sres = uop(uops, Ops.LOAD, dtypes.int32, (smem.index(ofs),))
     self.assertEqual(_test_uops_result(dtypes.int32, uops, sres), 42)
 
-@unittest.skipUnless(getenv("PTX"), "This only tests assembly backends")
+@unittest.skipUnless(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "This only tests assembly backends")
 class TestAssembly(unittest.TestCase):
   def test_bitshift_left(self):
     g1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)

test/test_uops_stats.py

@@ -8,6 +8,7 @@ from tinygrad.uop.ops import Ops, UOp
 from tinygrad.dtype import dtypes
 from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 from tinygrad.device import Device
+from tinygrad.renderer.ptx import PTXRenderer
 
 def flops_mem(uops, ignore_indexing=False):
   est = Estimates.from_uops(uops, ignore_indexing)
@@ -158,7 +159,7 @@ class TestUOpsStats(unittest.TestCase):
     self.assertEqual(flops_mem(uops), flops_mem(uops_fma))
 
 N = 64
-@unittest.skipIf(getenv("PTX"), "wrong in PTX") # maybe?
+@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "wrong in PTX") # maybe?
 class TestStatsOptimized(unittest.TestCase):
   @classmethod
   def setUpClass(cls):