OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Revert "[Inductor] support masked vectorization for the tail_loop for float64 datatype (#163316)"

Browse Source 
This reverts commit e9d8973427.

Reverted https://github.com/pytorch/pytorch/pull/163316 on behalf of https://github.com/clee2000 due to seems to have broken some no_gpu tests? test/inductor/test_cpu_repro.py::CPUReproTests::test_double_reduction_vec [GH job link](https://github.com/pytorch/pytorch/actions/runs/18689033019/job/53290772740) [HUD commit link](e9d8973427) ([comment](https://github.com/pytorch/pytorch/pull/163316#issuecomment-3428210509))
This commit is contained in:
PyTorch MergeBot 2025-10-21 17:44:42 +00:00
parent 78bf6186f2
commit 6c4412f72b
2 changed files with 0 additions and 65 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
test/inductor/test_cpu_repro.py
View File

@ -4810,22 +4810,6 @@ class CPUReproTests(TestCase):
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
# Tail vectorization case
x = torch.randn((22, 22), dtype=torch.double)
torch._dynamo.reset()
metrics.reset()
with torch.no_grad():
expected = fn(x)
compiled_fn = torch.compile(fn)
actual, code = run_and_get_cpp_code(compiled_fn, x)
self.assertEqual(expected, actual)
# 1 generated vec kernel
self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
# Check that both main and tail loops are vectorized
FileCheck().check_count(
"at::vec::VectorizedN<double,2>::loadu", 2, exactly=True
).run(code)
def test_double_reduction_vec(self):
def fn(x):
return x.sum(dim=1)
@ -4835,22 +4819,6 @@ class CPUReproTests(TestCase):
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
# Tail vectorization case
x = torch.randn((22, 22), dtype=torch.double)
torch._dynamo.reset()
metrics.reset()
with torch.no_grad():
expected = fn(x)
compiled_fn = torch.compile(fn)
actual, code = run_and_get_cpp_code(compiled_fn, x)
self.assertEqual(expected, actual)
# 1 generated vec kernel
self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
# Check that both main and tail loops are vectorized
FileCheck().check_count(
"at::vec::VectorizedN<double,2>::loadu", 2, exactly=True
).run(code)
def test_convert_fp32_to_double_vec(self):
def fn(x):
return x.to(torch.double)
@ -4860,22 +4828,6 @@ class CPUReproTests(TestCase):
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
# Tail vectorization case
x = torch.randn(22, 22)
torch._dynamo.reset()
metrics.reset()
with torch.no_grad():
expected = fn(x)
compiled_fn = torch.compile(fn)
actual, code = run_and_get_cpp_code(compiled_fn, x)
self.assertEqual(expected, actual)
# 1 generated vec kernel
self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
# Check that both main and tail loops are vectorized
FileCheck().check_count(
"at::vec::convert<double,2,float,1>", 2, exactly=True
).run(code)
def test_convert_double_to_fp32_vec(self):
def fn(x):
return x.to(torch.float32)
@ -4885,22 +4837,6 @@ class CPUReproTests(TestCase):
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
# Tail vectorization case
x = torch.randn((22, 22), dtype=torch.double)
torch._dynamo.reset()
metrics.reset()
with torch.no_grad():
expected = fn(x)
compiled_fn = torch.compile(fn)
actual, code = run_and_get_cpp_code(compiled_fn, x)
self.assertEqual(expected, actual)
# 1 generated vec kernel
self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
# Check that both main and tail loops are vectorized
FileCheck().check_count(
"at::vec::convert<float,1,double,2>", 2, exactly=True
).run(code)
def test_no_redundant_to_dtypes_between_fused_scheduler_node(self):
# https://github.com/pytorch/pytorch/issues/115260
p0 = torch.tensor([1.0879], dtype=torch.float16)

1
torch/_inductor/codegen/cpp.py
View File

@ -159,7 +159,6 @@ VECTORIZABLE_DTYPES: list[torch.dtype] = [
]
MASKED_VECTORIZABLE_DTYPES: list[torch.dtype] = [
torch.float64,
torch.float,
torch.bfloat16,
torch.float16,