prevent UB in long decomp and more emulated tests (#14447)

This commit is contained in:
2026-01-30 19:38:41 -05:00
committed by GitHub
parent 3204f94454
commit e575dd8275
5 changed files with 64 additions and 16 deletions
-3
View File
@@ -771,9 +771,6 @@ jobs:
run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
- name: Run TRANSCENDENTAL math
run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
- name: Test dtype with emulated long
if: matrix.backend != 'lvp'
run: EMULATED_DTYPES=long python3 -m pytest -n=auto test/test_dtype.py test/test_dtype_alu.py
- name: Run process replay tests
uses: ./.github/actions/process-replay
+1 -1
View File
@@ -1004,7 +1004,7 @@ def _get_runner(inst_bytes: bytes):
canonical_name = f"{_op_name(inst).lower()}_{base.to_bytes(size, 'little').hex()}"
sink = sink.replace(arg=KernelInfo(name=canonical_name)).rtag(1)
with Context(NOOPT=1, CHECK_OOB=0, TUPLE_ORDER=0):
with Context(NOOPT=1, CHECK_OOB=0, TUPLE_ORDER=0, EMULATED_DTYPES=""):
runner = get_runner('CPU', sink)
_canonical_runner_cache.append((base, mask, size, runner))
return runner, True
+27 -7
View File
@@ -1,4 +1,4 @@
import unittest, math
import contextlib, unittest, math
import numpy as np
import torch
from typing import Any, List
@@ -7,7 +7,8 @@ from tinygrad.helpers import getenv, DEBUG, CI, EMULATED_DTYPES
from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8, _to_np_dtype, _to_torch_dtype, truncate
from tinygrad.renderer.ptx import PTXRenderer
from tinygrad.renderer.nir import NIRRenderer
from tinygrad import Device, Tensor, dtypes
from tinygrad import Context, Device, Tensor, dtypes
from tinygrad.uop import Ops
from hypothesis import given, settings, strategies as strat
from test.helpers import rand_for_dtype
from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX
@@ -336,18 +337,37 @@ class TestUint16DType(TestDType):
class TestInt32DType(TestDType): DTYPE = dtypes.int32
class TestUint32DType(TestDType): DTYPE = dtypes.uint32
class TestInt64DType(TestDType):
DTYPE = dtypes.int64
class TestInt64DType(TestDType): DTYPE = dtypes.int64
@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
class TestEmulatedInt64DType(TestInt64DType):
@classmethod
def setUpClass(cls): cls.DATA = rand_for_dtype(cls.DTYPE, 10)
def setUpClass(cls):
cls.stack = contextlib.ExitStack()
cls.stack.enter_context(Context(EMULATED_DTYPES="long"))
cls.DATA = rand_for_dtype(cls.DTYPE, 10)
@classmethod
def tearDownClass(cls): cls.stack.close()
class TestUint64DType(TestDType):
@classmethod
def setUpClass(cls): cls.DATA = rand_for_dtype(cls.DTYPE, 10)
DTYPE = dtypes.uint64
def test_uint64_load(self):
assert Tensor(2**64 - 1, dtype=dtypes.uint64).numpy() == 2**64 - 1
@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
class TestEmulatedUInt64DType(TestUint64DType):
@classmethod
def setUpClass(cls):
cls.stack = contextlib.ExitStack()
cls.stack.enter_context(Context(EMULATED_DTYPES="long"))
cls.DATA = rand_for_dtype(cls.DTYPE, 10)
@classmethod
def tearDownClass(cls): cls.stack.close()
class TestBoolDType(TestDType): DTYPE = dtypes.bool
class TestBFloat16Type(TestDType): DTYPE = dtypes.bfloat16
+30 -1
View File
@@ -1,5 +1,5 @@
import unittest, operator, math
from tinygrad import Tensor, dtypes, Device
from tinygrad import Context, Tensor, dtypes, Device
from tinygrad.dtype import DType, truncate
from tinygrad.helpers import CI, getenv
from tinygrad.tensor import _to_np_dtype
@@ -7,6 +7,7 @@ from tinygrad.device import is_dtype_supported
from tinygrad.runtime.ops_python import from_storage_scalar
from tinygrad.renderer.ptx import PTXRenderer
from tinygrad.renderer.nir import NIRRenderer
from tinygrad.uop import Ops
import numpy as np
import pytest
from hypothesis import assume, given, strategies as strat, settings
@@ -165,9 +166,16 @@ class TestDTypeALU(unittest.TestCase):
@given(ht.uint32, ht.uint32, strat.sampled_from(integer_binary_operations))
def test_uint32(self, a, b, op): universal_test(a, b, dtypes.uint32, op)
@unittest.skipUnless(is_dtype_supported(dtypes.uint64), f"no uint64 on {Device.DEFAULT}")
@given(ht.uint64, ht.uint64, strat.sampled_from(integer_binary_operations))
def test_uint64(self, a, b, op): universal_test(a, b, dtypes.uint64, op)
@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
@given(ht.uint64, ht.uint64, strat.sampled_from(integer_binary_operations))
@Context(EMULATED_DTYPES="long")
def test_emulated_uint64(self, a, b, op): universal_test(a, b, dtypes.uint64, op)
@given(ht.int8, ht.int8, strat.sampled_from(integer_binary_operations))
def test_int8(self, a, b, op): universal_test(a, b, dtypes.int8, op)
@@ -177,9 +185,16 @@ class TestDTypeALU(unittest.TestCase):
@given(ht.int32, ht.int32, strat.sampled_from(integer_binary_operations))
def test_int32(self, a, b, op): universal_test(a, b, dtypes.int32, op)
@unittest.skipUnless(is_dtype_supported(dtypes.int64), f"no int64 on {Device.DEFAULT}")
@given(ht.int64, ht.int64, strat.sampled_from(integer_binary_operations))
def test_int64(self, a, b, op): universal_test(a, b, dtypes.int64, op)
@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
@given(ht.int64, ht.int64, strat.sampled_from(integer_binary_operations))
@Context(EMULATED_DTYPES="long")
def test_emulated_int64(self, a, b, op): universal_test(a, b, dtypes.int64, op)
@given(ht.uint8, strat.sampled_from(integer_unary_operations))
def test_uint8_unary(self, a, op): universal_test_unary(a, dtypes.uint8, op)
@@ -191,9 +206,16 @@ class TestDTypeALU(unittest.TestCase):
@given(ht.uint32, strat.sampled_from(integer_unary_operations))
def test_uint32_unary(self, a, op): universal_test_unary(a, dtypes.uint32, op)
@unittest.skipUnless(is_dtype_supported(dtypes.uint64), f"no uint64 on {Device.DEFAULT}")
@given(ht.uint64, strat.sampled_from(integer_unary_operations))
def test_uint64_unary(self, a, op): universal_test_unary(a, dtypes.uint64, op)
@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
@given(ht.uint64, strat.sampled_from(integer_unary_operations))
@Context(EMULATED_DTYPES="long")
def test_emulated_uint64_unary(self, a, op): universal_test_unary(a, dtypes.uint64, op)
@given(ht.int8, strat.sampled_from(integer_unary_operations))
def test_int8_unary(self, a, op): universal_test_unary(a, dtypes.int8, op)
@@ -203,9 +225,16 @@ class TestDTypeALU(unittest.TestCase):
@given(ht.int32, strat.sampled_from(integer_unary_operations))
def test_int32_unary(self, a, op): universal_test_unary(a, dtypes.int32, op)
@unittest.skipUnless(is_dtype_supported(dtypes.int64), f"no int64 on {Device.DEFAULT}")
@given(ht.int64, strat.sampled_from(integer_unary_operations))
def test_int64_unary(self, a, op): universal_test_unary(a, dtypes.int64, op)
@unittest.skipUnless(Ops.SHL in Device[Device.DEFAULT].renderer.code_for_op, "long decomp requires bitshift")
@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
@given(ht.int64, strat.sampled_from(integer_unary_operations))
@Context(EMULATED_DTYPES="long")
def test_emulated_int64_unary(self, a, op): universal_test_unary(a, dtypes.int64, op)
@given(ht.bool, ht.bool, strat.sampled_from(((operator.add, operator.add), (operator.mul, operator.mul))))
def test_bool(self, a, b, op): universal_test(a, b, dtypes.bool, op)
+6 -4
View File
@@ -353,8 +353,9 @@ def l2i(op: Ops, dt: DType, *uops:UOp):
case Ops.IDIV | Ops.MOD:
# TAOCP Algorithm 4.3.1D could be faster here, but must be parameterized over the width of b
if dt == dtypes.int:
a0, a1 = (a_neg:=a1 < zero).where((n:=l2i(Ops.NEG, dt, a0, a1))[0], a0).bitcast(dtypes.uint), a_neg.where(n[1], a1).bitcast(dtypes.uint)
b0, b1 = (b_neg:=b1 < zero).where((n:=l2i(Ops.NEG, dt, b0, b1))[0], b0).bitcast(dtypes.uint), b_neg.where(n[1], b1).bitcast(dtypes.uint)
ua0, ua1, ub0, ub1 = a0.bitcast(dtypes.uint), a1.bitcast(dtypes.uint), b0.bitcast(dtypes.uint), b1.bitcast(dtypes.uint)
a0, a1 = (a_neg:=a1 < zero).where((n:=l2i(Ops.NEG, dtypes.uint, ua0, ua1))[0], ua0), a_neg.where(n[1], ua1)
b0, b1 = (b_neg:=b1 < zero).where((n:=l2i(Ops.NEG, dtypes.uint, ub0, ub1))[0], ub0), b_neg.where(n[1], ub1)
q, r = (z:=UOp.const(dtypes.uint, 0), z), (z, z)
for i in range(63, -1, -1):
r = l2i(Ops.SHL, dtypes.uint, *r, UOp.const(dtypes.uint, 1), z)
@@ -364,8 +365,9 @@ def l2i(op: Ops, dt: DType, *uops:UOp):
q = ((q[0] | cond.cast(dtypes.uint) << (i % 32), q[1]) if i < 32 else (q[0], q[1] | cond.cast(dtypes.uint) << (i % 32)))
r = l2i(Ops.WHERE, dtypes.uint, cond, *diff, *r)
if dt == dtypes.int:
nq, nr = l2i(Ops.NEG, dt, q0:=q[0].bitcast(dt), q1:=q[1].bitcast(dt)), l2i(Ops.NEG, dt, r0:=r[0].bitcast(dt), r1:=r[1].bitcast(dt))
return (a_neg.where(nr[0], r0), a_neg.where(nr[1], r1)) if op == Ops.MOD else ((a_neg^b_neg).where(nq[0], q0), (a_neg^b_neg).where(nq[1], q1))
(nq0, nq1), (nr0, nr1) = l2i(Ops.BITCAST, dt, *l2i(Ops.NEG, dtypes.uint, *q)), l2i(Ops.BITCAST, dt, *l2i(Ops.NEG, dtypes.uint, *r))
(q0, q1), (r0, r1) = l2i(Ops.BITCAST, dt, *q), l2i(Ops.BITCAST, dt, *r)
return (a_neg.where(nr0, r0), a_neg.where(nr1, r1)) if op == Ops.MOD else ((a_neg^b_neg).where(nq0, q0), (a_neg^b_neg).where(nq1, q1))
return (r[0].bitcast(dt), r[1].bitcast(dt)) if op == Ops.MOD else (q[0].bitcast(dt), q[1].bitcast(dt))
case Ops.CMPLT: return (a1 < b1) | ((a1.eq(b1)) & (a0.bitcast(dtypes.uint) < b0.bitcast(dtypes.uint)))
case Ops.CMPEQ: return a0.eq(b0) & a1.eq(b1)