Google Git
Sign in
android / platform / external / pytorch / a8d6afb511a69687bbb2b7e88a3cf67917e1697e / . / test / inductor / test_cpu_repro.py
blob: 34d908d7eba05e3a2314eaec3a87c0b1af0478a9 [file] [log] [blame]
# Owner(s): ["oncall: cpu inductor"]
import contextlib
import copy
import functools
import itertools
import math
import platform
import sys
import unittest
from typing import Callable
from unittest.mock import patch
import torch
from torch import nn
from torch._C import FileCheck
from torch._dynamo.testing import rand_strided
from torch._dynamo.utils import same
from torch._inductor import config, cpu_vec_isa, metrics, test_operators
from torch._inductor.codegen.cpp import CppOverrides, CppVecOverrides
from torch._inductor.compile_fx import (
compile_fx,
compile_fx_inner,
complex_memory_overlap,
)
from torch._inductor.graph import GraphLowering
from torch._inductor.utils import timed
from torch._prims_common import is_float_dtype
from torch.fx.experimental.proxy_tensor import make_fx
from torch.nn import functional as F
from torch.testing._internal.common_utils import (
instantiate_parametrized_tests,
IS_FBCODE,
IS_MACOS,
parametrize,
skipIfRocm,
slowTest,
)
from torch.utils._python_dispatch import TorchDispatchMode
try:
try:
from . import test_torchinductor
except ImportError:
import test_torchinductor
except unittest.SkipTest:
if __name__ == "__main__":
sys.exit(0)
raise
vec_dtypes = test_torchinductor.vec_dtypes
_lowp_fp_dtypes = (
torch.bfloat16,
torch.float16,
)
run_and_get_cpp_code = test_torchinductor.run_and_get_cpp_code
TestCase = test_torchinductor.TestCase
aten = torch.ops.aten
check_model = test_torchinductor.check_model
requires_vectorization = unittest.skipUnless(
cpu_vec_isa.valid_vec_isa_list(), "Does not support vectorization"
)
def check_metrics_vec_kernel_count(num_expected_vec_kernels):
if cpu_vec_isa.valid_vec_isa_list():
assert metrics.generated_cpp_vec_kernel_count == num_expected_vec_kernels
@contextlib.contextmanager
def set_num_threads(num_threads):
orig_num_threads = torch.get_num_threads()
torch.set_num_threads(num_threads)
yield
torch.set_num_threads(orig_num_threads)
class LstmModule(torch.nn.Module):
def __init__(
self,
input_size,
hidden_size,
num_layers,
bias=True,
bidirectional=False,
batch_first=False,
):
super().__init__()
self.lstm = torch.nn.LSTM(
input_size=input_size,
hidden_size=hidden_size,
num_layers=num_layers,
bias=bias,
bidirectional=bidirectional,
batch_first=batch_first,
)
def forward(self, x, h=None):
x, h = self.lstm(x, h)
return x, h
@instantiate_parametrized_tests
class CPUReproTests(TestCase):
common = check_model
@skipIfRocm
def test_conv_stride_constraints(self):
for fmt in [torch.contiguous_format, torch.channels_last]:
# TorchDispatch doesn't work in our cuda invocation for some reason
m = torch.nn.Conv2d(5, 6, [3, 3])
def fn(inp, weight):
return (
F.conv2d(
inp, weight, None, m.stride, m.padding, m.dilation, m.groups
),
)
inp = torch.randn([2, 5, 16, 16])
inps = [inp, m.weight.to(memory_format=fmt)]
fn_fx = make_fx(fn)(*inps)
fn_compiled = compile_fx_inner(fn_fx, inps)
test_self = self
conv_seen = False
class RecordFunctions(TorchDispatchMode):
def __torch_dispatch__(self, func, types, args=(), kwargs=None):
kwargs = kwargs if kwargs else {}
if func == torch.ops.aten.convolution.default:
# For CPU and mkldnn enable, we always using channles last
nonlocal fmt
if (
torch.backends.mkldnn.enabled
and torch.backends.mkldnn.is_available()
):
fmt = torch.channels_last
test_self.assertTrue(args[0].is_contiguous(memory_format=fmt))
test_self.assertTrue(args[1].is_contiguous(memory_format=fmt))
nonlocal conv_seen
conv_seen = True
return func(*args, **kwargs)
with RecordFunctions():
out = fn_compiled(inps)
self.assertTrue(conv_seen)
@patch("torch.cuda.is_available", lambda: False)
def test_conv2d_bn_mixed_dtype(self):
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv = torch.nn.Conv2d(
3,
16,
kernel_size=3,
stride=1,
padding=1,
bias=False,
dtype=torch.bfloat16,
)
self.bn = torch.nn.BatchNorm2d(
16, eps=0.001, momentum=0.1, affine=True, track_running_stats=True
)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
return x
v = torch.randn(1, 3, 64, 64, dtype=torch.bfloat16)
mod = Model().eval()
with torch.no_grad():
self.common(
mod,
(v,),
)
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
def test_conv2d_packed(self):
options = itertools.product([[3, 56, 56]], [True, False], [0, (0,)])
for x_shape, mode_train, padding in options:
mod = torch.nn.Sequential(
torch.nn.Conv2d(3, 64, 3, 3, padding=padding)
).train(mode=mode_train)
v = torch.randn(x_shape, dtype=torch.float32)
with torch.no_grad():
self.common(
mod,
(v,),
)
@patch("torch.cuda.is_available", lambda: False)
def test_conv2d_autocast(self):
v = torch.randn(1, 3, 28, 18, dtype=torch.float32)
mod = torch.nn.Sequential(torch.nn.Conv2d(3, 64, 3, 3)).eval()
with torch.no_grad(), torch.cpu.amp.autocast():
self.common(
mod,
(v,),
)
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
def test_unsupported_conv_transpose(self):
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv_transpose = torch.nn.ConvTranspose2d(
3, 6, 3, stride=1, padding=1, output_padding=1
)
def forward(self, input_tensor):
x = self.conv_transpose(input_tensor)
output = torch.tanh(x)
return output
input = torch.randn(1, 3, 28, 28)
m = Model().eval()
with torch.no_grad():
compiled_m = torch.compile(m)
with self.assertRaisesRegex(
RuntimeError,
"output padding must be smaller than either stride or dilation",
):
compiled_m(input)
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
def test_conv_used_from_multiple_places(self):
class M(torch.nn.Module):
def __init__(self, conv_in_channel, conv_out_channel) -> None:
super().__init__()
self.conv = torch.nn.Conv2d(conv_in_channel, conv_out_channel, (3, 3))
def forward(self, x):
res = self.conv(x)
res = F.relu(res)
res = self.conv(res)
return res
with torch.no_grad():
mod = M(3, 3).eval()
x = torch.randn(1, 3, 224, 224)
self.common(
mod,
(x,),
)
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
def test_linear_used_from_multiple_places(self):
class M(torch.nn.Module):
def __init__(self, in_channel, out_channel) -> None:
super().__init__()
self.linear = torch.nn.Linear(in_channel, out_channel)
def forward(self, x):
res = self.linear(x)
res = F.relu(res)
res = self.linear(res)
return res
dtypes = []
if torch.ops.mkldnn._is_mkldnn_bf16_supported():
dtypes.append(torch.bfloat16)
if torch.ops.mkldnn._is_mkldnn_fp16_supported():
dtypes.append(torch.float16)
for dtype in dtypes:
with torch.no_grad():
m = M(224, 224).to(dtype).eval()
m_opt = torch.compile(m)
x = torch.randn(224, 224, dtype=dtype)
m_opt(x)
self.assertEqual(m(x), m_opt(x))
@config.patch(implicit_fallbacks=True)
def test_multihead_attention_cpu(self):
def fn(
q,
k,
v,
embed_dim,
num_heads,
qkv_weight,
qkv_bias,
proj_weight,
proj_bias,
mask,
need_weights,
):
return torch._native_multi_head_attention(
q,
k,
v,
embed_dim,
num_heads,
qkv_weight,
qkv_bias,
proj_weight,
proj_bias,
mask,
need_weights,
)
B = 1
T = 3
embed_dim = 6
num_heads = 2
q = torch.randn([B, T, embed_dim])
k = torch.randn([B, T, embed_dim])
v = torch.randn([B, T, embed_dim])
qkv_weight = torch.randn([3 * embed_dim, embed_dim])
qkv_bias = torch.randn([3 * embed_dim])
proj_weight = torch.randn([3 * embed_dim, embed_dim])
proj_bias = torch.randn([3 * embed_dim])
mask = None
need_weights = False
inps = [
q,
k,
v,
embed_dim,
num_heads,
qkv_weight,
qkv_bias,
proj_weight,
proj_bias,
mask,
need_weights,
]
self.common(fn, inps)
@config.patch(freezing=True)
def test_module_buffer_mutation(self):
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.foo = torch.nn.Buffer(torch.rand((3, 10)))
def forward(self, x):
lx = [x, x.clone(), x.clone()]
y = []
for i in range(3):
y.append(lx[i] + self.foo[i])
return torch.cat(y, 1)
with torch.no_grad():
example_inputs = (torch.rand(1, 10),)
self.common(Model(), example_inputs)
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
def test_linear_packed(self):
dtypes = []
if torch.ops.mkldnn._is_mkldnn_bf16_supported():
dtypes.append(torch.bfloat16)
if torch.ops.mkldnn._is_mkldnn_fp16_supported():
dtypes.append(torch.float16)
options = itertools.product(
[[2, 3, 10], [2, 10], [10], [2, 0]], [3, 0], [True, False], dtypes
)
for input_shape, out_dim, bias, dtype in options:
mod = torch.nn.Sequential(
torch.nn.Linear(input_shape[-1], out_dim, bias=bias)
).eval()
v = torch.randn(input_shape)
with torch.no_grad():
self.common(
mod.to(dtype),
(v.to(dtype),),
)
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
def test_conv_transpose2d_packed_cpu(self):
options = itertools.product([[1, 3, 28, 28], [3, 28, 28]], [0, (0,)])
for x_shape, padding in options:
mod = torch.nn.Sequential(
torch.nn.ConvTranspose2d(3, 64, 3, 3, padding=padding)
).eval()
v = torch.randn(x_shape, dtype=torch.float32)
with torch.no_grad():
self.common(
mod,
(v,),
)
@torch._dynamo.config.patch(
{"dynamic_shapes": True, "assume_static_by_default": False}
)
def test_full_boolean_dynamic_shape(self):
def fn(n):
x = torch.full((1024,), n >= 1024)
return x, x + 1
self.common(fn, (1024,))
self.common(fn, (1023,))
@config.patch(freezing=True)
@unittest.skipIf(not torch._C._has_mkldnn, "MKLDNN is not enabled")
@torch._dynamo.config.patch(dynamic_shapes=True)
@torch._dynamo.config.patch(assume_static_by_default=False)
def test_conv_in_channel_1_dynamic_shapes(self):
class M(torch.nn.Module):
def __init__(self, in_channel, out_channel) -> None:
super().__init__()
self.conv = torch.nn.Conv2d(in_channel, out_channel, 3)
def forward(self, x):
res = self.conv(x)
res = F.relu(res)
return res
# test the case where the channels dim of the input is 1
# Reproducer from the maml_omniglot model in Torchbench
in_channel = 1
out_channel = 3
amp_enabled_configs = [False]
if torch.ops.mkldnn._is_mkldnn_bf16_supported():
# When amp is enabled here, the input to Conv is a FlexibleLayout.
# While it's disabled, the input is a FixedLayout.
amp_enabled_configs.append(True)
for amp_enabled in amp_enabled_configs:
mod = M(in_channel, out_channel).eval()
v = torch.randn(5, in_channel, 15, 15)
with torch.no_grad(), torch.cpu.amp.autocast(enabled=amp_enabled):
self.common(
mod,
(v,),
)
@unittest.skipIf(not torch._C._has_mkldnn, "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
@torch._dynamo.config.patch(dynamic_shapes=True)
@torch._dynamo.config.patch(assume_static_by_default=False)
@torch._dynamo.config.patch(allow_rnn=True)
@config.patch(freezing=True)
def _test_lstm_packed(self, params_dict, change_input_sizes=False):
from torch._dynamo.utils import counters
for (
unbatched,
input_size,
hidden_size,
num_layers,
bidirectional,
bias,
empty_state,
batch_first,
batch_size,
seq_len,
) in itertools.product(*list(params_dict.values())):
dtypes = [torch.float]
if torch.ops.mkldnn._is_mkldnn_bf16_supported():
dtypes.append(torch.bfloat16)
if torch.ops.mkldnn._is_mkldnn_fp16_supported():
dtypes.append(torch.float16)
for dtype in dtypes:
counters.clear()
num_directions = 2 if bidirectional else 1
seq_len_var = seq_len + 3
if unbatched:
v = torch.randn(seq_len, input_size)
v_var = torch.randn(seq_len_var, input_size)
h = torch.randn(num_layers * num_directions, hidden_size)
c = torch.randn(num_layers * num_directions, hidden_size)
else:
if batch_first:
v = torch.randn(batch_size, seq_len, input_size)
v_var = torch.randn(batch_size, seq_len_var, input_size)
else:
v = torch.randn(seq_len, batch_size, input_size)
v_var = torch.randn(seq_len_var, batch_size, input_size)
h = torch.randn(
num_layers * num_directions, batch_size, hidden_size
)
c = torch.randn(
num_layers * num_directions, batch_size, hidden_size
)
mod = LstmModule(
input_size,
hidden_size,
num_layers,
bias,
bidirectional,
batch_first,
).eval()
maybe_autocast = (
torch.cpu.amp.autocast()
if dtype == torch.bfloat16
else contextlib.nullcontext()
)
with torch.no_grad(), maybe_autocast:
inps = [v]
if not empty_state:
inps.append((h, c))
fn_opt = torch._dynamo.optimize("inductor")(mod)
_, code = run_and_get_cpp_code(fn_opt, *inps)
# Check that _flat_weights are not functional_tensor, otherwise
# deepcopy will fail during recompilation.
fn_opt_copy = copy.deepcopy(fn_opt)
_flat_weights = fn_opt_copy.lstm._flat_weights
for _flat_weight in _flat_weights:
self.assertFalse(torch._is_functional_tensor(_flat_weight))
self.assertTrue("aten.mkldnn_rnn_layer" in code)
self.assertEqual(fn_opt(*inps), mod(*inps))
self.assertEqual(
counters["inductor"]["pattern_matcher_count"],
num_layers * num_directions
+ 2, # num of mkldnn_rnn_layer call + 2 view call on the concatenated hy, cy.
)
# Change input sizes
if change_input_sizes:
inps_var = [v_var]
self.assertEqual(fn_opt(*inps_var), mod(*inps_var))
@slowTest
def test_lstm_packed(self):
params_dict = {
"unbatched": [True, False],
"input_size": [1, 2],
"hidden_size": [2],
"num_layers": [1, 2],
"bidirectional": [False, True],
"bias": [False, True],
"empty_state": [False, True],
"batch_first": [True, False],
"batch_size": [1, 2],
"seq_len": [1, 2],
}
self._test_lstm_packed(params_dict)
def test_lstm_packed_change_input_sizes_cpu(self):
params_dict = {
"unbatched": [False],
"input_size": [2],
"hidden_size": [5],
"num_layers": [3],
"bidirectional": [True],
"bias": [True],
"empty_state": [False],
"batch_first": [False],
"batch_size": [2],
"seq_len": [3],
}
self._test_lstm_packed(params_dict, change_input_sizes=True)
@torch._dynamo.config.patch(dynamic_shapes=True)
@torch._dynamo.config.patch(assume_static_by_default=False)
@torch._dynamo.config.patch(allow_rnn=True)
def test_pack_padded_sequence_lstm(self):
embedding_dim = 12
hidden_dim = 10
batch_size = 24
num_layers = 1
bidirectional = True
num_direc = 2
max_lens = 96
sent = torch.randn(batch_size, max_lens, embedding_dim)
hid_0 = torch.rand(num_layers * num_direc, batch_size, hidden_dim)
hid_1 = torch.randn(num_layers * num_direc, batch_size, hidden_dim)
sent_lens = torch.Tensor(
[1, 2, 3, 4, 5, 1, 3, 2, 96, 5, 3, 1, 1, 2, 1, 2, 3, 6, 1, 2, 4, 6, 2, 1]
)
assert sent_lens.shape[0] == batch_size
assert sent_lens.max().item() == max_lens
hidden_0 = hid_0.clone().requires_grad_(False)
hidden_1 = hid_1.clone().requires_grad_(False)
embeds = torch.nn.utils.rnn.pack_padded_sequence(
sent, sent_lens, batch_first=True, enforce_sorted=False
)
mod = LstmModule(
embedding_dim,
hidden_dim,
num_layers=num_layers,
bias=True,
bidirectional=bidirectional,
batch_first=True,
).eval()
with torch.no_grad():
inps = [embeds, (hidden_0, hidden_1)]
fn_opt = torch._dynamo.optimize("inductor")(mod)
_, code = run_and_get_cpp_code(fn_opt, *inps)
# This case is unsupported
self.assertFalse("torch.ops.mkldnn._lstm" in code)
self.assertEqual(fn_opt(*inps), mod(*inps))
@patch("torch.cuda.is_available", lambda: False)
def test_conv_transpose2d_has_output_size_input(self):
# https://github.com/pytorch/pytorch/issues/100344.
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv_transpose = torch.nn.ConvTranspose2d(
in_channels=3, out_channels=1, kernel_size=3, stride=1, padding=1
)
def forward(self, x):
return self.conv_transpose(x, output_size=(10, 10))
mod = M().eval()
v = torch.randn(1, 3, 10, 10, dtype=torch.float32)
with torch.no_grad():
self.common(
mod,
(v,),
)
def test_pad_with_nan_value(self):
# https://github.com/pytorch/pytorch/issues/100988.
class Model(torch.nn.Module):
def forward(self, x):
x = F.pad(x, (1, 1, 1, 1), value=float("nan"))
return x
mod = Model().eval()
v = torch.randn(1, 3, 10, 10, dtype=torch.float32)
with torch.no_grad():
self.common(
mod,
(v,),
)
def test_masked_fill_with_inf_or_nan_value(self):
def fn(value, mask):
y1 = torch.masked_fill(value, mask, float("inf"))
y2 = torch.masked_fill(value, mask, float("-inf"))
y3 = torch.masked_fill(value, mask, float("nan"))
return y1, y2, y3
value = torch.randn((2, 17))
mask = torch.randint(0, 1, size=(2, 17), dtype=torch.uint8).to(torch.bool)
with torch.no_grad():
self.common(
fn,
(value, mask),
)
def test_relu_with_inf_value(self):
# https://github.com/pytorch/pytorch/issues/117544.
def fn(out):
out = torch.sinh(input=out)
out = torch.relu(input=out)
return out
x = torch.Tensor([-572373.5000, 755109.1250, 330995.5625])
with torch.no_grad():
self.common(
fn,
(x,),
)
def test_acosh_with_negative_large_input(self):
# https://github.com/pytorch/pytorch/issues/118267.
def fn(input):
out = torch.acosh(input)
return out
x = torch.Tensor(
[
[
-8493.9854,
431654.1250,
71741.5859,
608234.5000,
-103814.7500,
-699397.0000,
-910685.8125,
-832737.1875,
875343.5000,
]
]
).repeat(3, 9)
for dtype in [torch.float32, torch.bfloat16, torch.double]:
with torch.no_grad():
torch._dynamo.reset()
metrics.reset()
_x = x.to(dtype)
self.common(
fn,
(_x,),
)
@config.patch(implicit_fallbacks=True)
def test_repeat_interleave(self):
def fn(y):
return torch.repeat_interleave(y, 2, output_size=8)
a = torch.tensor([[1, 2], [3, 4]])
self.common(
fn,
(a,),
)
def test_inplace_squeeze_needed(self):
mod = torch.nn.Sequential(
torch.nn.Linear(10, 10),
torch.nn.LayerNorm(10),
torch.nn.ReLU(),
).eval()
def fn(x):
return mod(x)
v = torch.randn(10)
# TODO: OMP parallel reduction order is not deterministic.
# Hence, the accurarcy might vary up and down. For short term,
# we increase the tolerance and will fix it later by using
# aten parallel.
self.common(fn, (v,), atol=5e-1, rtol=5e-1)
def test_cat_mul(self):
# https://github.com/pytorch/pytorch/issues/93365
def fn(p0, p1):
y1 = torch.cat([p0, p1], dim=0)
y2 = torch.mul(y1, y1)
return y1, y2
p0 = torch.randn(3, 4)
p1 = torch.randn(3, 4)
self.common(fn, (p0, p1))
def test_pow_cos(self):
# https://github.com/pytorch/pytorch/issues/98149
def fn(x):
t = x.pow(5)
return torch.cos(t)
x = torch.tensor([4], dtype=torch.uint8)
self.common(fn, (x,))
def test_reduce_with_masked(self):
# https://github.com/pytorch/pytorch/issues/96484
def fn(a, b):
a = torch.nn.functional.pad(a, (0, -1))
c = a + b
return c.min(0).values
a = torch.randn([2])
b = torch.randn([2])
self.common(fn, (a, b))
def test_scalar_sign_with_min(self):
# https://github.com/pytorch/pytorch/issues/101340
def fn(a):
t1 = torch.tanh(a)
t2 = torch.sign(t1)
return torch.min(t1, t2)
a = torch.randn(1, 3)
self.common(fn, (a,))
def test_index_propagation_issue_102065(self):
def fn(x):
x = torch.arange(x.numel())
return (x.unsqueeze(0) - x.unsqueeze(1)) ** 2
self.common(
fn,
(torch.randn(8),),
)
def test_ModularIndexing_range_issue_103133(self):
def fn(q, k):
einsum = torch.einsum("bcxd,bcyd->bcxy", (q, k))
constant_pad_nd = torch.ops.aten.constant_pad_nd.default(
einsum, [0, 0, 0, 1], 0.0
)
view = torch.ops.aten.view.default(constant_pad_nd, [12, 1, 512, 513])
y = view.new_zeros((12, 2, 256, 513))
y[:, :-1, :, 256:] = view[:, :, :256, :257]
return y
self.common(
fn,
(
torch.empty_strided((12, 1, 512, 64), (64, 196608, 768, 1)),
torch.empty_strided((12, 1, 512, 64), (64, 196608, 768, 1)),
),
)
@patch("torch.cuda.is_available", lambda: False)
def test_max_reduction_lowp_fp(self):
def fn(x):
return torch.ops.aten.max(x, 1, keepdim=True)[0].float()
for dtype in _lowp_fp_dtypes:
self.common(
fn,
(torch.randn(1, 32, 4, 4).to(dtype),),
)
@patch("torch.cuda.is_available", lambda: False)
def test_vec_transpose_lowp_fp(self):
for dtype in _lowp_fp_dtypes:
def fn(x):
return x.to(memory_format=torch.channels_last).to(dtype)
self.common(
fn,
(torch.randn(2, 3, 4, 4),),
)
def test_load_inf_bf16(self):
def fn1(x):
return torch.where(x > 0, x, math.inf)
def fn2(x):
return torch.where(x > 0, x, -math.inf)
for fn in [fn1, fn2]:
self.common(
fn,
(torch.randn(1, 3, 16, 16),),
)
@patch("torch.cuda.is_available", lambda: False)
def test_fp32_load_with_to_lowp_fp(self):
# From llama model.
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.cache_k = torch.zeros(8, 4, 2, 2)
def forward(self, x, xk):
bsz, seqlen, _ = x.shape
self.cache_k = self.cache_k.to(x)
self.cache_k[:bsz, 1 : 1 + seqlen] = xk
return self.cache_k
for dtype in _lowp_fp_dtypes:
ref_model = Model().eval()
opt_model = torch.compile()(Model().eval())
x = torch.randn(4, 2, 2).to(dtype)
xk = torch.randn(4, 2, 2, 2).to(dtype)
self.assertEqual(opt_model(x, xk), ref_model(x, xk))
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_sigmoid_with_reduction(self):
def fn(x):
x = torch.ops.aten.sigmoid.default(x)
return torch.ops.aten.mean.dim(x, [-1, -2], True)
x = torch.randn((1, 8, 8, 8))
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
def test_slice_scatter_default_end_value(self):
# From HF AllenaiLongformerBase.
def fn(query, key, window_overlap):
batch_size, seq_len, num_heads, head_dim = query.size()
assert (
seq_len % (window_overlap * 2) == 0
), f"Sequence length should be multiple of {window_overlap * 2}. Given {seq_len}"
chunks_count = torch.div(seq_len, window_overlap, rounding_mode="trunc") - 1
diagonal_chunked_attention_scores = key
diagonal_attention_scores = diagonal_chunked_attention_scores.new_zeros(
(
batch_size * num_heads,
chunks_count + 1,
window_overlap,
window_overlap * 2 + 1,
)
)
diagonal_attention_scores[
:, :3, :, window_overlap:
] = diagonal_chunked_attention_scores[
:, :, :window_overlap, : window_overlap + 1
]
return diagonal_attention_scores
self.common(
fn,
(
torch.randn(1, 1024, 12, 64),
torch.randn(12, 3, 512, 513),
256,
),
)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_to_uint8_rounding_method(self):
def fn(x):
return x.to(torch.uint8)
numerical_testsuit = [4.4, 4.5, 4.6, 5.5]
for numerical_number in numerical_testsuit:
x = torch.ones(17) * numerical_number
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
@requires_vectorization
def _test_decomposed_dequant_relu_quant_helper(self, dtype):
def fn(
x, scale, zero_point, use_dequant, use_quant, quant_min, quant_max, dtype
):
# For quantized_decomposed.dequantize_per_tensor
# Refer to torch/ao/quantization/fx/_decomposed.py
if use_dequant:
x = (x.to(torch.float32) - zero_point) * scale
x = torch.relu(x)
# For quantized_decomposed.quantize_per_tensor
# Refer to torch/ao/quantization/fx/_decomposed.py
if use_quant:
inv_scale = 1.0 / scale
x = torch.clamp(
torch.round(x * inv_scale) + zero_point, quant_min, quant_max
).to(dtype)
return x
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
use_dequant_list = [False, True]
use_quant_list = [False, True]
for use_dequant, use_quant in itertools.product(
use_dequant_list, use_quant_list
):
x = torch.clamp(
torch.randn((1, 7, 7, 9), dtype=torch.float32) * 100,
quant_min,
quant_max,
)
if use_dequant:
x = x.to(dtype)
zero_point = 100
scale = 0.01
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(
fn,
(
x,
scale,
zero_point,
use_dequant,
use_quant,
quant_min,
quant_max,
dtype,
),
)
check_metrics_vec_kernel_count(1)
@requires_vectorization
def test_decomposed_dequant_relu_quant_uint8(self):
self._test_decomposed_dequant_relu_quant_helper(torch.uint8)
@requires_vectorization
def test_decomposed_dequant_relu_quant_int8(self):
self._test_decomposed_dequant_relu_quant_helper(torch.int8)
def _test_dequant_quant_lowering_helper(self, dtype):
def fn(
x, scale, zero_point, use_dequant, use_quant, quant_min, quant_max, dtype
):
if use_dequant:
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
x, scale, zero_point, quant_min, quant_max, dtype
)
x = torch.relu(x)
if use_quant:
x = torch.ops.quantized_decomposed.quantize_per_tensor(
x, scale, zero_point, quant_min, quant_max, dtype
)
return x
use_dequant_list = [False, True]
use_quant_list = [False, True]
use_tensor_overload_list = [False, True]
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
for use_dequant, use_quant, use_tensor_overload in itertools.product(
use_dequant_list, use_quant_list, use_tensor_overload_list
):
x = torch.clamp(
torch.randn((1, 7, 7, 9), dtype=torch.float32) * 100,
quant_min,
quant_max,
)
if use_dequant:
x = x.to(dtype)
zero_point = 100
scale = 0.01
if use_tensor_overload:
zero_point = torch.tensor(zero_point, dtype=torch.int64)
scale = torch.tensor(scale)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(
fn,
(
x,
scale,
zero_point,
use_dequant,
use_quant,
quant_min,
quant_max,
dtype,
),
)
check_metrics_vec_kernel_count(1)
@requires_vectorization
def test_dequant_quant_lowering_uint8(self):
self._test_dequant_quant_lowering_helper(torch.uint8)
@requires_vectorization
def test_dequant_quant_lowering_int8(self):
self._test_dequant_quant_lowering_helper(torch.int8)
def _test_dequant_maxpool2d_lowering_helper(self, dtype):
def fn(x, scale, zero_point, quant_min, quant_max, dtype):
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
x, scale, zero_point, quant_min, quant_max, dtype
)
max_pool2d_with_indices_default = (
torch.ops.aten.max_pool2d_with_indices.default(
x, [2, 2], [2, 2], [1, 1]
)[0]
)
return max_pool2d_with_indices_default
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
use_tensor_overload_list = [False, True]
for use_tensor_overload in use_tensor_overload_list:
x = (
torch.clamp(
torch.randn((3, 16, 8, 8), dtype=torch.float32) * 100,
quant_min,
quant_max,
)
.to(dtype)
.contiguous(memory_format=torch.channels_last)
)
zero_point = 100
scale = 0.01
if use_tensor_overload:
zero_point = torch.tensor(zero_point, dtype=torch.int64)
scale = torch.tensor(scale)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x, scale, zero_point, quant_min, quant_max, dtype))
check_metrics_vec_kernel_count(1)
@requires_vectorization
def test_dequant_maxpool2d_lowering_uint8(self):
self._test_dequant_maxpool2d_lowering_helper(torch.uint8)
@requires_vectorization
def test_dequant_maxpool2d_lowering_int8(self):
self._test_dequant_maxpool2d_lowering_helper(torch.int8)
def _test_tile2d_load_decomposed_dequant_add_relu_quant_helper(self, dtype):
def fn(
x,
scale,
zero_point,
x2,
scale2,
zero_point2,
output_scale,
output_zero_point,
use_dequant,
use_dequant2,
use_quant,
quant_min,
quant_max,
dtype,
):
if use_dequant:
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
x, scale, zero_point, quant_min, quant_max, dtype
)
if use_dequant2:
x2 = torch.ops.quantized_decomposed.dequantize_per_tensor(
x2, scale2, zero_point2, quant_min, quant_max, dtype
)
temp = x + x2
y = torch.relu(temp)
if use_quant:
y = torch.ops.quantized_decomposed.quantize_per_tensor(
y, output_scale, output_zero_point, quant_min, quant_max, dtype
)
return y.contiguous()
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
use_dequant_list = [False, True]
use_dequant_list2 = [False, True]
use_quant_list = [False, True]
for use_dequant, use_dequant2, use_quant in itertools.product(
use_dequant_list, use_dequant_list2, use_quant_list
):
x = torch.clamp(
torch.randn((1, 1024, 14, 14), dtype=torch.float32) * 100,
quant_min,
quant_max,
).contiguous(memory_format=torch.channels_last)
x2 = torch.clamp(
torch.randn((1, 1024, 14, 14), dtype=torch.float32) * 100,
quant_min,
quant_max,
).contiguous(memory_format=torch.channels_last)
if use_dequant:
x = x.to(dtype).contiguous(memory_format=torch.channels_last)
if use_dequant2:
x2 = x2.to(dtype).contiguous(memory_format=torch.channels_last)
zero_point = 1
scale = 0.01
zero_point2 = 2
scale2 = 0.02
output_zero_point = 3
output_scale = 0.03
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(
fn,
(
x,
scale,
zero_point,
x2,
scale2,
zero_point2,
output_scale,
output_zero_point,
use_dequant,
use_dequant2,
use_quant,
quant_min,
quant_max,
dtype,
),
)
check_metrics_vec_kernel_count(2)
@requires_vectorization
def test_tile2d_load_decomposed_dequant_add_relu_quant_uint8(self):
self._test_tile2d_load_decomposed_dequant_add_relu_quant_helper(torch.uint8)
@requires_vectorization
def test_tile2d_load_decomposed_dequant_add_relu_quant_int8(self):
self._test_tile2d_load_decomposed_dequant_add_relu_quant_helper(torch.int8)
@requires_vectorization
def _test_per_tensor_fake_quant_helper(self, dtype):
def fn(input, scales, zero_points, quant_min, quant_max, dtype):
input = torch.ops.quantized_decomposed.quantize_per_tensor(
input, scales, zero_points, quant_min, quant_max, dtype
)
input = torch.ops.quantized_decomposed.dequantize_per_tensor(
input, scales, zero_points, quant_min, quant_max, dtype
)
return input
use_tensor_overload_list = [False, True]
for use_tensor_overload in use_tensor_overload_list:
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
x = torch.clamp(
torch.randn((1, 7, 7, 9), dtype=torch.float32) * 100,
quant_min,
quant_max,
)
zero_point = 100
scale = 0.01
if use_tensor_overload:
zero_point = torch.tensor(zero_point, dtype=torch.int64)
scale = torch.tensor(scale)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x, scale, zero_point, quant_min, quant_max, dtype))
assert metrics.generated_cpp_vec_kernel_count == 1
@requires_vectorization
def test_per_tensor_fake_quant_uint8(self):
self._test_per_tensor_fake_quant_helper(torch.uint8)
@requires_vectorization
def test_per_tensor_fake_quant_int8(self):
self._test_per_tensor_fake_quant_helper(torch.int8)
def _test_per_channel_fake_quant_helper(self, dtype, input_dtype=torch.float32):
def fn(input, scales, zero_points, axis, quant_min, quant_max, dtype):
input = torch.ops.quantized_decomposed.quantize_per_channel(
input, scales, zero_points, axis, quant_min, quant_max, dtype
)
input = torch.ops.quantized_decomposed.dequantize_per_channel(
input, scales, zero_points, axis, quant_min, quant_max, dtype
)
return input
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
x = torch.clamp(
torch.randn((1, 3, 224, 224), dtype=torch.float32) * 100,
quant_min,
quant_max,
)
if input_dtype != torch.float32:
x = x.to(dtype=input_dtype)
scales = torch.ones((3,))
zero_points = torch.zeros((3,))
axis = 1
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x, scales, zero_points, axis, quant_min, quant_max, dtype))
check_metrics_vec_kernel_count(1)
@requires_vectorization
def test_per_channel_fake_quant_uint8(self):
self._test_per_channel_fake_quant_helper(torch.uint8)
@requires_vectorization
def test_per_channel_fake_quant_module_uint8(self):
class Mod(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.scales = torch.ones((3,)).to(torch.float64)
self.zero_points = torch.zeros((3,)).to(torch.int64)
self.axis = 1
self.quant_min = 0
self.quant_max = 255
self.dtype = torch.uint8
def forward(self, input):
input = torch.ops.quantized_decomposed.quantize_per_channel(
input,
self.scales,
self.zero_points,
self.axis,
self.quant_min,
self.quant_max,
self.dtype,
)
input = torch.ops.quantized_decomposed.dequantize_per_channel(
input,
self.scales,
self.zero_points,
self.axis,
self.quant_min,
self.quant_max,
self.dtype,
)
return input
m = Mod().eval()
x = torch.clamp(
torch.randn((1, 3, 224, 224), dtype=torch.float32) * 100,
0,
255,
)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(m, (x,))
assert metrics.generated_cpp_vec_kernel_count == 1
@requires_vectorization
def test_per_channel_fake_quant_int8(self):
self._test_per_channel_fake_quant_helper(torch.int8)
@requires_vectorization
def test_per_channel_fake_quant_uint8_bf16_input(self):
self._test_per_channel_fake_quant_helper(
torch.uint8, input_dtype=torch.bfloat16
)
@requires_vectorization
def test_per_channel_fake_quant_int8_bf16_input(self):
self._test_per_channel_fake_quant_helper(torch.int8, input_dtype=torch.bfloat16)
def _test_non_contiguous_load_buf_quant_helper(self, dtype):
def fn(
x1,
x2,
groups,
quant_min,
quant_max,
dtype,
):
x = torch.cat((x1, x2), dim=1)
batchsize, num_channels, height, width = x.size()
channels_per_group = num_channels // groups
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
x, 1.0, 0, quant_min, quant_max, dtype
)
x = x.view(batchsize, groups, channels_per_group, height, width)
x = torch.ops.quantized_decomposed.quantize_per_tensor(
x, 1.0, 0, quant_min, quant_max, dtype
)
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
x, 1.0, 0, quant_min, quant_max, dtype
)
x = torch.transpose(x, 1, 2).contiguous()
x = x.view(batchsize, num_channels, height, width)
return x
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
x = torch.randint(0, 8, (1, 116, 28, 28), dtype=dtype).contiguous(
memory_format=torch.channels_last
)
x2 = torch.randint(0, 8, (1, 116, 28, 28), dtype=dtype).contiguous(
memory_format=torch.channels_last
)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(
fn,
(
x,
x2,
2,
quant_min,
quant_max,
dtype,
),
)
check_metrics_vec_kernel_count(2)
@requires_vectorization
def test_non_contiguous_load_buf_quant_uint8(self):
self._test_non_contiguous_load_buf_quant_helper(torch.uint8)
@requires_vectorization
def test_non_contiguous_load_buf_quant_int8(self):
self._test_non_contiguous_load_buf_quant_helper(torch.int8)
def _test_tile2d_store_channel_shuffle_cl_quant_output_helper(self, dtype):
def channel_shuffle(
x, groups, output_scale, output_zero_point, quant_min, quant_max, dtype
):
batchsize, num_channels, height, width = x.size()
channels_per_group = num_channels // groups
x = x.view(batchsize, groups, channels_per_group, height, width)
x = torch.transpose(x, 1, 2).contiguous()
x = x.view(batchsize, -1, height, width)
x = torch.ops.quantized_decomposed.quantize_per_tensor(
x, output_scale, output_zero_point, quant_min, quant_max, dtype
)
return x.contiguous(memory_format=torch.channels_last)
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
x = torch.randn(64, 58, 28, 28)
output_zero_point = 3
output_scale = 0.03
self.common(
channel_shuffle,
(x, 2, output_scale, output_zero_point, quant_min, quant_max, dtype),
)
check_metrics_vec_kernel_count(2)
@requires_vectorization
def test_tile2d_store_channel_shuffle_cl_quant_output_uint8(self):
self._test_tile2d_store_channel_shuffle_cl_quant_output_helper(torch.uint8)
@requires_vectorization
def test_tile2d_store_channel_shuffle_cl_quant_output_int8(self):
self._test_tile2d_store_channel_shuffle_cl_quant_output_helper(torch.int8)
def _test_dequant_relu_quant_dequant_relu_quant_lowering_helper(self, dtype):
def fn(
x,
scale,
zero_point,
scale2,
zero_point2,
scale3,
zero_point3,
quant_min,
quant_max,
dtype,
):
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
x, scale, zero_point, quant_min, quant_max, dtype
)
x = torch.relu(x)
x = torch.ops.quantized_decomposed.quantize_per_tensor(
x, scale2, zero_point2, quant_min, quant_max, dtype
)
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
x, scale2, zero_point2, quant_min, quant_max, dtype
)
x = torch.relu(x)
x = torch.ops.quantized_decomposed.quantize_per_tensor(
x, scale3, zero_point3, quant_min, quant_max, dtype
)
return x
assert dtype in [torch.uint8, torch.int8]
quant_min = 0 if dtype == torch.uint8 else -128
quant_max = 255 if dtype == torch.uint8 else 127
for use_tensor_overload in [True, False]:
x = torch.clamp(
torch.randn((1, 7, 7, 9), dtype=torch.float32) * 100,
quant_min,
quant_max,
).to(dtype)
zero_point_list = [100, 101, 102]
scale_list = [0.01, 0.02, 0.03]
if use_tensor_overload:
for i in range(len(zero_point_list)):
zero_point_list[i] = torch.tensor(
zero_point_list[i], dtype=torch.int64
)
scale_list[i] = torch.tensor(scale_list[i])
zero_point, zero_point2, zero_point3 = zero_point_list
scale, scale2, scale3 = scale_list
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(
fn,
(
x,
scale,
zero_point,
scale2,
zero_point2,
scale3,
zero_point3,
quant_min,
quant_max,
dtype,
),
rtol=1e-2,
atol=1e-2,
)
check_metrics_vec_kernel_count(1)
@requires_vectorization
def test_dequant_relu_quant_dequant_relu_quant_lowering_uint8(self):
self._test_dequant_relu_quant_dequant_relu_quant_lowering_helper(torch.uint8)
@requires_vectorization
def test_dequant_relu_quant_dequant_relu_quant_lowering_int8(self):
self._test_dequant_relu_quant_dequant_relu_quant_lowering_helper(torch.int8)
def test_inplace_add_alpha(self):
def fn(x, y):
aten.add_.Tensor(x, y, alpha=0.55)
return (x,)
x1 = torch.zeros(10)
x2 = torch.zeros(10)
x3 = torch.zeros(10)
y = torch.randn(10)
fn_fx = make_fx(fn)(x1, y)
fn_compiled = compile_fx_inner(fn_fx, [x1, y])
fn(x2, y)
fn_compiled([x3, y])
assert same(x2, x3)
def test_int_div(self):
def fn(x, y):
s3 = x.size(1)
a = torch.ones((1 + s3) // 2)
a += y
return a, s3
p0 = torch.randint(5, (1, 8))
p1 = torch.randn(1)
self.common(fn, (p0, p1))
def test_no_op_squeeze(self):
@torch._dynamo.optimize("inductor")
def forward(arg0_1):
return torch.ops.aten.squeeze.dim(arg0_1, 1)
x = torch.randn((10, 20))
self.common(forward, (x,))
def test_parallel_num_threads(self):
@torch._dynamo.optimize("inductor")
def fn(x1, x2):
return x1 + x2
x1 = torch.randn((10, 20))
x2 = torch.randn((10, 20))
with set_num_threads(1):
assert same(x1 + x2, fn(x1, x2))
with set_num_threads(4):
assert same(x1 + x2, fn(x1, x2))
@patch("torch.cuda.is_available", lambda: False)
def test_timed_cpu_only(self):
timed(lambda: torch.randn(10), ())
def test_complex_memory_overlap(self):
dense = torch.zeros(64, 32)
self.assertFalse(complex_memory_overlap(dense))
self.assertFalse(complex_memory_overlap(dense.t()))
strided = dense.split(4, dim=1)
self.assertFalse(complex_memory_overlap(strided[0]))
self.assertFalse(complex_memory_overlap(strided[0].t()))
unsqueezed = dense.unsqueeze(1)
self.assertFalse(complex_memory_overlap(unsqueezed))
self.assertFalse(complex_memory_overlap(unsqueezed.permute(1, 2, 0)))
gathered = dense.index_select(0, torch.IntTensor([1, 0, 1]))
self.assertFalse(complex_memory_overlap(gathered))
self.assertFalse(complex_memory_overlap(gathered.t()))
@requires_vectorization
def test_vec_dynamic_shapes(self):
def fn(x):
return torch.softmax(x, -1)
value = torch.randn((2, 10))
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (value,))
@unittest.skipIf(IS_FBCODE, "Not yet runnable in fbcode")
@unittest.skipIf(
platform.machine() != "x86_64" or not cpu_vec_isa.valid_vec_isa_list(),
"Does not support vectorization or not x86_64 machine",
)
@patch("torch.cuda.is_available", lambda: False)
def test_auto_simd(self):
vec_amx = cpu_vec_isa.supported_vec_isa_list[0]
vec_avx512 = cpu_vec_isa.supported_vec_isa_list[1]
vec_avx2 = cpu_vec_isa.supported_vec_isa_list[2]
self.assertTrue(vec_amx.bit_width() == 512)
self.assertTrue(vec_amx.nelements() == 16)
self.assertTrue(vec_amx.nelements(torch.bfloat16) == 32)
self.assertTrue(vec_avx512.bit_width() == 512)
self.assertTrue(vec_avx2.bit_width() == 256)
self.assertTrue(vec_avx512.nelements() == 16)
self.assertTrue(vec_avx2.nelements() == 8)
self.assertTrue(vec_avx512.nelements(torch.bfloat16) == 32)
self.assertTrue(vec_avx2.nelements(torch.bfloat16) == 16)
with config.patch({"cpp.simdlen": None}):
isa = cpu_vec_isa.pick_vec_isa()
if vec_amx in cpu_vec_isa.valid_vec_isa_list():
self.assertTrue(isa == vec_amx)
elif vec_avx512 in cpu_vec_isa.valid_vec_isa_list():
self.assertTrue(isa == vec_avx512)
else:
self.assertTrue(isa == vec_avx2)
with config.patch({"cpp.simdlen": 0}):
isa = cpu_vec_isa.pick_vec_isa()
self.assertFalse(isa)
with config.patch({"cpp.simdlen": 1}):
isa = cpu_vec_isa.pick_vec_isa()
self.assertFalse(isa)
with config.patch({"cpp.simdlen": 257}):
isa = cpu_vec_isa.pick_vec_isa()
self.assertFalse(isa)
with config.patch({"cpp.simdlen": 513}):
isa_list = cpu_vec_isa.valid_vec_isa_list()
if vec_avx512 in isa_list:
self.assertFalse(isa)
with config.patch({"cpp.simdlen": 512}):
isa_list = cpu_vec_isa.valid_vec_isa_list()
isa = cpu_vec_isa.pick_vec_isa()
if vec_amx in isa_list:
self.assertTrue(isa == vec_amx)
elif vec_avx512 in isa_list:
self.assertTrue(isa == vec_avx512)
with config.patch({"cpp.simdlen": 256}):
isa_list = cpu_vec_isa.valid_vec_isa_list()
if vec_avx2 in isa_list:
isa = cpu_vec_isa.pick_vec_isa()
self.assertTrue(isa == vec_avx2)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_masked_fill_softmax(self):
def fn(value, mask):
mask = mask.to(torch.bool)
x = torch.masked_fill(value, mask, -33.0)
return torch.softmax(x, -1)
for dtype in vec_dtypes:
value = torch.randn((2, 17), dtype=dtype)
mask = torch.randint(0, 1, size=(2, 17), dtype=torch.uint8)
with config.patch({"cpp.simdlen": None}):
for cpp_wrapper_flag in [True, False]:
with config.patch({"cpp_wrapper": cpp_wrapper_flag}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (value, mask))
assert metrics.generated_cpp_vec_kernel_count >= 1
def test_channels_last_view_as_complex(self):
# https://github.com/pytorch/pytorch/issues/122448#issuecomment-2046169554
def reduce_example(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
"""Applies the rotary embedding to the query and key tensors."""
x_out = torch.view_as_complex(torch.stack([x.float(), y.float()], dim=-1))
return x_out
args = [torch.randn(1, 1, 1, 128), torch.randn(1, 1, 1, 128)]
expected = reduce_example(*args)
actual = torch.compile(reduce_example, fullgraph=True)(*args)
self.assertEqual(expected, actual)
def test_load_same_bool_tensor_twice(self):
@torch._dynamo.optimize("inductor")
def fn(a, b):
x = torch.masked_fill(a, b, -33.0)
y = torch.masked_fill(a, b, -33.0)
return x, y
value = torch.randn((2, 17))
mask = torch.randint(0, 1, size=(2, 17), dtype=torch.uint8).to(torch.bool)
fn(value, mask)
def test_cpu_vec_cosim(self):
cpp_vec_op_list = []
cpp_op_list = []
for k, v in CppVecOverrides.__dict__.items():
if isinstance(v, staticmethod):
cpp_vec_op_list.append(k)
for k, v in CppOverrides.__dict__.items():
if isinstance(v, staticmethod):
cpp_op_list.append(k)
diff = [
"airy_ai",
"bessel_j0",
"bessel_j1",
"bessel_y0",
"bessel_y1",
"modified_bessel_i0",
"modified_bessel_i1",
"modified_bessel_k0",
"modified_bessel_k1",
"scaled_modified_bessel_k0",
"scaled_modified_bessel_k1",
"spherical_bessel_j0",
"i1",
"i1e",
"ndtr",
"ndtri",
"log_ndtr",
"erfcx",
"gammainc",
"gammaincc",
"igamma",
"igammac",
"polygamma",
"zeta",
"shifted_chebyshev_polynomial_u",
"chebyshev_polynomial_u",
"chebyshev_polynomial_t",
"shifted_chebyshev_polynomial_w",
"chebyshev_polynomial_w",
"shifted_chebyshev_polynomial_t",
"chebyshev_polynomial_v",
"shifted_chebyshev_polynomial_v",
"hermite_polynomial_he",
"laguerre_polynomial_l",
"hermite_polynomial_h",
"legendre_polynomial_p",
"constant",
"index_expr",
"signbit",
"isinf",
"frexp",
"mod",
"masked",
"randn",
"isnan",
"rand",
"randint64",
"logical_and",
"logical_not",
"logical_or",
"logical_xor",
"bitwise_and",
"bitwise_left_shift",
"bitwise_not",
"bitwise_right_shift",
"bitwise_or",
"bitwise_xor",
"to_dtype_bitcast",
]
union = {*cpp_vec_op_list, *diff}
self.assertTrue(
set(cpp_op_list).issubset(union), f"unexpected: {set(cpp_op_list) - union}"
)
def test_atomic_add_lowp_fp(self):
def fn(test_args):
res = torch.gather(**test_args)
return res
for dtype in _lowp_fp_dtypes:
input_tensor_for_ref = torch.tensor(
[[3.0, -5.0]], dtype=dtype, requires_grad=True
)
input_tensor_for_opt = torch.tensor(
[[3.0, -5.0]], dtype=dtype, requires_grad=True
)
test_args_for_ref = {
"input": input_tensor_for_ref,
"dim": 1,
"index": torch.tensor([[1]]),
}
test_args_for_opt = {
"input": input_tensor_for_opt,
"dim": 1,
"index": torch.tensor([[1]]),
}
opt_fn = torch.compile(fn)
ref_fwd = fn(test_args_for_ref)
res_fwd = opt_fn(test_args_for_opt)
self.assertEqual(res_fwd, ref_fwd)
torch.manual_seed(1)
bwd_tensor_for_ref = torch.randn(ref_fwd.shape, dtype=dtype)
torch.manual_seed(1)
bwd_tensor_for_opt = torch.randn(res_fwd.shape, dtype=dtype)
self.assertEqual(bwd_tensor_for_ref, bwd_tensor_for_opt)
ref_fwd.backward(bwd_tensor_for_ref)
res_fwd.backward(bwd_tensor_for_opt)
ref_grad = test_args_for_ref["input"].grad
res_grad = test_args_for_opt["input"].grad
self.assertEqual(ref_grad, res_grad)
def test_meta_device(self):
@torch.compile(fullgraph=True)
def fn():
x = torch.ops.aten.empty.memory_format(
[1024, 128, 128],
dtype=torch.float16,
device="meta",
pin_memory=False,
)
return x.sin() + 1
self.assertEqual(fn().shape, [1024, 128, 128])
def test_decomposed_fake_quant_per_channel(self):
def fq(input, scales, zero_points, axis, quant_min, quant_max):
res = torch.fake_quantize_per_channel_affine(
input, scales, zero_points, axis, quant_min, quant_max
)
return res
def qdq(input, scales, zero_points, axis, quant_min, quant_max):
res = torch.ops.quantized_decomposed.fake_quant_per_channel(
input, scales, zero_points, axis, quant_min, quant_max
)
return res
def run_eager_aten_fake_quant(
input, scales, zero_points, axis, quant_min, quant_max
):
input.grad = None
res = fq(input, scales, zero_points, axis, quant_min, quant_max)
res.sum().backward()
return res, input.grad
def run_eager_decomposed_fake_quant(
input, scales, zero_points, axis, quant_min, quant_max
):
input.grad = None
res = qdq(input, scales, zero_points, axis, quant_min, quant_max)
res.sum().backward()
return res, input.grad
def run_compile_decomposed_fake_quant(
input, scales, zero_points, axis, quant_min, quant_max
):
input.grad = None
compiled_qdq = torch.compile(qdq)
res = compiled_qdq(input, scales, zero_points, axis, quant_min, quant_max)
res.sum().backward()
return res, input.grad
input = torch.randn(2, 3, 224, 224)
input[1, 2, 3, 4] = 257
input.requires_grad_()
scales = torch.ones((3,))
zero_points = torch.zeros((3,))
axis = 1
quant_min = -128
quant_max = 127
aten_input = copy.deepcopy(input)
compiler_input = copy.deepcopy(input)
res_aten_eager, input_grad_aten_eager = run_eager_aten_fake_quant(
aten_input, scales, zero_points, axis, quant_min, quant_max
)
res_decomp_eager, input_grad_decomp_eager = run_eager_decomposed_fake_quant(
input, scales, zero_points, axis, quant_min, quant_max
)
res, input_grad = run_compile_decomposed_fake_quant(
compiler_input, scales, zero_points, axis, quant_min, quant_max
)
self.assertEqual(res_aten_eager, res)
self.assertEqual(res_decomp_eager, res)
self.assertEqual(input_grad_aten_eager, input_grad)
self.assertEqual(input_grad_decomp_eager, input_grad)
self.assertEqual(input_grad[1, 2, 3, 4], torch.tensor(0.0))
# For forward and backward kernel
check_metrics_vec_kernel_count(2)
@requires_vectorization
def test_ops_masked_with_bool_input(self):
x = torch.zeros(129, dtype=torch.bool)
size = [2, 3]
res_aten_eager = torch.constant_pad_nd(x, size)
cfn = torch.compile(torch.constant_pad_nd)
res = cfn(x, size)
self.assertEqual(res_aten_eager, res)
check_metrics_vec_kernel_count(1)
def test_bitwise_right_shift(self):
x = torch.randint(-1, 0, (1, 1, 1), device="cpu", dtype=torch.int64)
bit_num = 31
res_aten_eager = torch.bitwise_right_shift(x, bit_num)
cfn = torch.compile(torch.bitwise_right_shift)
res = cfn(x, bit_num)
self.assertEqual(res_aten_eager, res)
def test_view_dtype(self):
def f(x):
return x.view(torch.int32) >> 2
input = torch.ones(16, 16)
res_aten_eager = f(input)
cfn = torch.compile(f)
res = cfn(input)
self.assertEqual(res_aten_eager, res)
@patch("torch.cuda.is_available", lambda: False)
def test_scatter_using_atomic_add(self):
def fn(a, dim, index, b):
return aten.scatter(a, dim, index, b, reduce="add")
inps = (
torch.randn(5, 29, 13),
2,
torch.tensor([[[3, 5, 7, 9]]]),
torch.randn(1, 1, 10),
)
def _internal_check(
_fn,
_inps,
_target_code_check=None,
_target_code_check_not=None,
):
torch._dynamo.reset()
metrics.reset()
_fn_opt = torch.compile()(_fn)
_, code = run_and_get_cpp_code(_fn_opt, *inps)
if _target_code_check:
FileCheck().check(_target_code_check).run(code)
if _target_code_check_not:
FileCheck().check_not(_target_code_check_not).run(code)
# Verify that the output isn't empty
FileCheck().check("Output code:").run(code)
self.assertEqual(
_fn(*_inps),
_fn_opt(*_inps),
)
with config.patch({"cpp.fallback_scatter_reduce_sum": False}):
_internal_check(fn, inps, "atomic_add")
with config.patch({"cpp.fallback_scatter_reduce_sum": True}):
_internal_check(fn, inps, "aten.scatter_reduce_")
if "ATen parallel backend: OpenMP" in torch.__config__.parallel_info():
with set_num_threads(1):
# When running with a single thread, we expect the aten.scatter will go
# into the cpp backend codegen instead of a fallback to aten.scatter_reduce_.
# Avoid the inductor cache so we don't serve an entry compiled above.
with config.patch(
{"fx_graph_cache": False, "fx_graph_remote_cache": False}
):
_internal_check(
fn, inps, _target_code_check_not="aten.scatter_reduce_"
)
with config.patch({"cpp.dynamic_threads": True}), set_num_threads(1):
_internal_check(fn, inps, "aten.scatter_reduce_")
@patch("torch.cuda.is_available", lambda: False)
@requires_vectorization
@torch._inductor.config.patch({"cpp.fallback_scatter_reduce_sum": False})
def test_scatter_using_atomic_add_vec(self):
def fn(a, dim, index, b):
return aten.scatter(a, dim, index, b, reduce="add")
inps = (
torch.zeros(1, 1, 25),
2,
torch.tensor([[[3, 5, 7, 9] * 5]]),
torch.ones(1, 1, 25),
)
torch._dynamo.reset()
metrics.reset()
self.common(fn, inps)
assert metrics.generated_cpp_vec_kernel_count == 2
with set_num_threads(1), config.patch(
{"fx_graph_cache": False, "fx_graph_remote_cache": False}
):
torch._dynamo.reset()
metrics.reset()
self.common(fn, inps)
assert metrics.generated_cpp_vec_kernel_count == 2
@unittest.skipIf(IS_FBCODE, "Not yet runnable in fbcode")
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_new_vec_op_cpu_only(self):
def fn(x):
return torch.log1p(torch.expm1(torch.erf(x)))
for dtype in vec_dtypes:
torch.manual_seed(0)
x = torch.randn((2, 9), dtype=dtype)
x[0, 0] = torch.nan
x[1, -1] = torch.nan
tol = 1e-2 if dtype == torch.bfloat16 else 1e-4
with config.patch({"cpp.simdlen": None}):
for cpp_wrapper_flag in [True, False]:
with config.patch({"cpp_wrapper": cpp_wrapper_flag}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_vec_cpu_only_for_all_available_isa(self):
def fn(x):
return torch.sin(torch.cos(torch.erf(x)))
x = torch.randn((2, 9))
x[0, 0] = torch.nan
x[1, -1] = torch.nan
bit_widths = [isa._bit_width for isa in cpu_vec_isa.valid_vec_isa_list()] + [
None
]
for item in bit_widths:
with config.patch({"cpp.simdlen": item}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
@slowTest
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
@config.patch("cpp.enable_tiling_heuristics", False)
def test__adaptive_avg_pool2d(self):
def wrap_fn(oh, ow):
def fn(x):
return torch._adaptive_avg_pool2d(x, (oh, ow))
return fn
bit_widths = [isa._bit_width for isa in cpu_vec_isa.valid_vec_isa_list()]
ih = [16, 65]
iw = ih
oh = ih
ow = ih
for _ih, _iw, _oh, _ow, _simd_len, dtype in itertools.product(
ih, iw, oh, ow, bit_widths, vec_dtypes
):
x = torch.randn(2, 3, _ih, _iw, dtype=dtype).to(
memory_format=torch.channels_last
)
_fn = wrap_fn(_oh, _ow)
with config.patch({"cpp.simdlen": _simd_len}):
torch._dynamo.reset()
metrics.reset()
self.common(_fn, (x,))
check_metrics_vec_kernel_count(1)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_vec_logical(self):
def wrap_fn1(op: Callable):
def fn(x: torch.Tensor):
return torch.where(op(x), 1.0, 0.0)
return fn
def wrap_fn2(op: Callable):
def fn(x: torch.Tensor, y: torch.Tensor):
return torch.where(op(x, y), 1.0, 0.0)
return fn
for dtype in vec_dtypes:
x = torch.randn(64, dtype=dtype)
y = torch.randn(64, dtype=dtype)
logical_fns = [
torch.logical_and,
torch.logical_not,
torch.logical_or,
torch.logical_xor,
]
for logical_fn in logical_fns:
torch._dynamo.reset()
metrics.reset()
if logical_fn == torch.logical_not:
_fn = wrap_fn1(logical_fn)
_args = (x,)
else:
_fn = wrap_fn2(logical_fn)
_args = (x, y)
self.common(_fn, _args)
check_metrics_vec_kernel_count(1)
@requires_vectorization
def test_vec_bitwise(self):
for dtype in [
torch.bool,
torch.uint8,
torch.int8,
torch.int32,
torch.int64,
]:
x = torch.randn(64, dtype=torch.float32)
y = torch.randn(64, dtype=torch.float32)
if dtype == torch.bool:
x = x > 0
y = y > 0
else:
x = x.to(dtype)
y = y.to(dtype)
bitwise_fns = [
torch.bitwise_and,
torch.bitwise_not,
torch.bitwise_or,
torch.bitwise_xor,
torch.bitwise_left_shift,
torch.bitwise_right_shift,
]
for bitwise_fn in bitwise_fns:
if (
bitwise_fn
in [
torch.bitwise_left_shift,
torch.bitwise_right_shift,
]
and dtype == torch.bool
):
# Eager doesn't support bool
# https://pytorch.org/docs/stable/generated/torch.bitwise_left_shift.html
continue
torch._dynamo.reset()
metrics.reset()
if bitwise_fn == torch.bitwise_not:
_args = (x,)
else:
_args = (x, y)
self.common(bitwise_fn, _args)
check_metrics_vec_kernel_count(1)
@requires_vectorization
def test_vec_randn(self):
funcs = [torch.randn, torch.rand, torch.randint]
float_dtypes = [
torch.bfloat16,
torch.float16,
torch.float32,
]
int_dtypes = [
torch.int8,
torch.uint8,
torch.int32,
torch.int64,
]
dtypes = float_dtypes + int_dtypes
for rand_func, dtype in itertools.product(funcs, dtypes):
if (rand_func == torch.randint and dtype not in int_dtypes) or (
rand_func != torch.randint and dtype not in float_dtypes
):
# Skip the invalid combination
continue
with config.patch(
{"fx_graph_cache": False, "fx_graph_remote_cache": False}
):
torch._dynamo.reset()
metrics.reset()
def func(seed):
torch.manual_seed(seed)
if rand_func == torch.randint:
return rand_func(0, 64, (64,), dtype=dtype)
else:
return rand_func(64, dtype=dtype)
cfn = torch.compile(func)
# Check the result is deterministic with mauanl seed
self.assertEqual(cfn(2024), cfn(2024))
check_metrics_vec_kernel_count(1)
res_vec = cfn(2024)
torch._dynamo.reset()
metrics.reset()
with config.patch({"cpp.simdlen": 0}):
res_scalar = torch.compile(func)(2024)
# Check the same result between scalar and vec
self.assertEqual(res_vec, res_scalar)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_vec_compare_op_cpu_only(self):
def fn(x):
y1 = torch.eq(x, 1.0)
x = torch.where(y1, x, -x)
y2 = torch.ne(x, 0.0)
x = torch.where(y2, x, -x)
y3 = torch.lt(x, 5.0)
x = torch.where(y3, x, x - 1.0)
y4 = torch.gt(x, -2.0)
x = torch.where(y4, x, x + 1.0)
y5 = torch.le(x, 8.0)
x = torch.where(y5, x, x - 1.0)
y6 = torch.ge(x, -3.0)
x = torch.where(y6, x, x + 1.0)
y7 = x == 1.0
x = torch.where(y7, x, -x)
y8 = x != 0.0
x = torch.where(y8, x, -x)
y9 = x < 5.0
x = torch.where(y9, x, x - 1.0)
y10 = x > -2.0
x = torch.where(y10, x, x + 1.0)
y11 = x <= 8.0
x = torch.where(y11, x, x - 1.0)
y12 = x >= -3.0
x = torch.where(y12, x, x + 1.0)
return x
for dtype in vec_dtypes:
x = torch.randn((2, 9), dtype=dtype)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
assert (
metrics.generated_kernel_count
- metrics.generated_cpp_vec_kernel_count
) == 0
@requires_vectorization
def test_vec_remainder(self):
for dtype in [
torch.int8,
torch.uint8,
torch.int32,
torch.int64,
torch.bfloat16,
torch.float16,
torch.float32,
torch.float64,
]:
if is_float_dtype(dtype):
x = torch.randn(64, dtype=dtype)
y = torch.randn(64, dtype=dtype)
else:
lower = 1 if dtype == torch.uint8 else -100
x = torch.randint(lower, 100, (64,), dtype=dtype)
y = torch.randint(lower, 100, (64,), dtype=dtype)
y = torch.where(
y == torch.zeros_like(y),
torch.ones_like(y),
y,
)
torch._dynamo.reset()
metrics.reset()
_args = (x, y)
self.common(torch.remainder, _args)
check_metrics_vec_kernel_count(1)
def test_skip_cpp_codegen(self):
with config.patch({"disable_cpp_codegen": True}):
inps = torch.ones([20]), torch.rand([20])
def f(x, y):
return x + y + torch.tensor(1)
f_opt = torch.compile()(f)
_, code = run_and_get_cpp_code(f_opt, inps[0], inps[1])
FileCheck().check_not("void kernel").run(code)
self.assertEqual(
f(*inps),
f_opt(*inps),
)
# constant needs to be propagated on fallback
def f(x):
return x[torch.tensor(1) :] * 2
f_opt = torch.compile()(f)
_, code = run_and_get_cpp_code(f_opt, inps[0])
FileCheck().check_not("void kernel").run(code)
self.assertEqual(f_opt(inps[0]), f(inps[0]))
class Model(torch.nn.Module):
def __init__(
self,
):
super().__init__()
def forward(self, v1: torch.Tensor):
vx = v1.min(dim=1).values
v2 = torch.randn_like(vx)
return v2
model = Model()
x = torch.rand(10, 3, 0)
model_f = torch.compile()(model)
self.assertEqual(model(x), model_f(x))
def test_redundant_to_node_elimination_lowp_fp(self):
def fn(x, y):
res = x + y
res = torch.mean(res)
return res
for dtype in _lowp_fp_dtypes:
x = torch.randn((2, 9), dtype=dtype)
y = torch.randn((2, 9), dtype=dtype)
for torch_compile_debug in [True, False]:
with config.patch(
{"trace.enabled": torch_compile_debug, "cpp.simdlen": None}
):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x, y))
check_metrics_vec_kernel_count(1)
def test_do_not_insert_to_dtype_for_memory_copy_only_kernel(self):
def fn(x):
res = x.clone()
return res
x = torch.randn((100, 100), dtype=torch.bfloat16)
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
assert metrics.cpp_to_dtype_count == 0
check_metrics_vec_kernel_count(1)
def test_insert_to_dtype_count(self):
def fn(x):
res = x.relu()
return res
x = torch.randn((100, 100), dtype=torch.bfloat16)
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
assert metrics.cpp_to_dtype_count == 2
check_metrics_vec_kernel_count(1)
def test_memory_copy_with_fusion(self):
def fn(x):
res = x.relu()
x.copy_(res)
return (res,)
x = torch.randn((100, 100), dtype=torch.bfloat16)
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
assert metrics.cpp_to_dtype_count == 2
check_metrics_vec_kernel_count(1)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_maxpool2d_cpu_only(self):
for dtype in vec_dtypes:
input = torch.randn(26, 32, 112, 112, dtype=dtype).to(
memory_format=torch.channels_last
)
maxpool = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
def func(x):
return maxpool(x)
with patch.object(config.cpp, "simdlen", None):
torch._dynamo.reset()
metrics.reset()
self.common(func, (input,))
check_metrics_vec_kernel_count(1)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
@config.patch("cpp.enable_tiling_heuristics", False)
def test_maxpool2d_with_pre_loop_collapse_cpu_only(self):
x1 = torch.randn(2, 3, 20, 20).to(memory_format=torch.channels_last)
x2 = torch.randn(2, 3, 20, 20).to(memory_format=torch.channels_last)
maxpool = torch.nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
def func(x1, x2):
y = x1 + x2
return maxpool(y)
with patch.object(config.cpp, "simdlen", None):
torch._dynamo.reset()
metrics.reset()
self.common(func, (x1, x2))
check_metrics_vec_kernel_count(2)
def test_randint_symint_input(self):
# https://github.com/pytorch/pytorch/issues/122405
@torch.compile(fullgraph=True)
def get_traj_idx(lengths: torch.Tensor, num_slices: int) -> torch.Tensor:
return torch.randint(lengths.shape[0], (num_slices,), device=lengths.device)
lengths = torch.zeros(10, dtype=torch.long)
get_traj_idx(lengths, num_slices=4)
lengths = torch.zeros(11, dtype=torch.long)
get_traj_idx(lengths, num_slices=4)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_sign_cpu_only(self):
def fn(x):
return torch.sign(x)
for dtype in vec_dtypes:
x = torch.randn((2, 9), dtype=dtype)
x[0, 0] = torch.nan
x[1, -1] = torch.nan
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_reduction_cpu_only(self):
def fn(x):
return torch.argmax(x, -1)
for dtype in vec_dtypes:
x = torch.randn((10, 10), dtype=dtype)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
assert metrics.generated_cpp_vec_kernel_count == 1
@config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
def test_outer_loop_fusion(self):
def fn(x):
max = torch.amax(x, dim=-1, keepdim=True)
return x - max
x = torch.randn(4, 12, 1023, 1022)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
self.assertEqual(
len(metrics.cpp_outer_loop_fused_inner_counts),
1,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].inner_kernel_number,
2,
)
@config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
def test_local_buffer_in_outer_loop_fusion(self):
def fn(x):
max = torch.nn.functional.softmax(x, dim=-1)
return x - max
x = torch.randn(4, 12, 1023, 1022)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
self.assertEqual(
len(metrics.cpp_outer_loop_fused_inner_counts),
1,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].inner_kernel_number,
3,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].local_buffer_number,
1,
)
# Check the number of global buffer allocation
torch._dynamo.reset()
metrics.reset()
_, code = run_and_get_cpp_code(
torch._dynamo.optimize("inductor")(fn),
x,
)
self.assertEqual(code.count("empty_strided_cpu("), 3)
@config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
def test_two_local_buffers_in_outer_loop_fusion(self):
def fn(x):
softmax = torch.nn.functional.softmax(x, dim=-1)
sum = torch.sum(softmax, dim=-1)
sum_broadcast = torch.broadcast_to(
sum.unsqueeze(-1), [*(sum.size()[0:3]), 256]
)
sum_exp = torch.exp(sum_broadcast)
sum2 = torch.sum(sum_exp, dim=-1)
sub = sum_exp - sum2.unsqueeze(-1)
return x[:, :, :, 0:256] - sub
x = torch.randn(4, 12, 1023, 1022)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
self.assertEqual(
len(metrics.cpp_outer_loop_fused_inner_counts),
1,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].inner_kernel_number,
5,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].local_buffer_number,
2,
)
@config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
def test_share_local_buffers_in_outer_loop_fusion(self):
def fn(x):
max = torch.nn.functional.softmax(x, dim=-1)
max = torch.nn.functional.softmax(max, dim=-1)
return x - max
x = torch.randn(4, 12, 1023, 1022)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
self.assertEqual(
len(metrics.cpp_outer_loop_fused_inner_counts),
1,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].inner_kernel_number,
5,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].local_buffer_number,
1, # 2 global bufs share 1 local buf
)
@config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
def test_two_local_buffers_in_outer_loop_fusion_case2(self):
# exp and exp2 should be replaced by local buffer
# since exp will be used after exp2, exp2 can't share the same
# local buffer as exp
def fn(x):
a_max = torch.amax(x, -1, keepdim=True)
exp = torch.exp(x - a_max)
sum = torch.sum(exp, -1, keepdim=True)
exp2 = torch.exp(exp - sum)
sum2 = torch.sum(exp2, -1, keepdim=True)
sub = exp2 - sum2
sub2 = exp - sub
return sub2
x = torch.randn(4, 12, 1023, 1022)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
self.assertEqual(
len(metrics.cpp_outer_loop_fused_inner_counts),
1,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].inner_kernel_number,
4,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].local_buffer_number,
2,
)
@config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
def test_local_buffer_with_line_reuse(self):
# Test Global buffer which is inplace buffer and replaced by local buffer
def fn(x, y):
z = torch.matmul(x, y)
a_max = torch.amax(x, -1, keepdim=True)
# Previous is a inplace buffer and now is a local buffer
exp = torch.exp((z - a_max) / z)
sum = torch.sum(exp, -1, keepdim=True)
return exp - sum
inputs = [torch.rand(4, 32), torch.rand(32, 32)]
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, inputs)
self.assertEqual(
len(metrics.cpp_outer_loop_fused_inner_counts),
1,
)
self.assertEqual(
metrics.cpp_outer_loop_fused_inner_counts[0].local_buffer_number,
1,
)
def test_argmin(self):
def fn(x):
return torch.argmin(x, -1)
for dtype in vec_dtypes:
x = torch.randn((10, 10), dtype=dtype)
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
assert metrics.generated_cpp_vec_kernel_count == 1
def test_argmax_argmin_with_nan_value(self):
def fn(x):
return torch.argmax(x)
def fn2(x):
return torch.argmin(x)
inputs = [
torch.Tensor([-755832.1250, 100]),
torch.Tensor([-755832.1250, 100, 200]),
torch.Tensor([100, -755832.1250]),
torch.Tensor([100, 200, -755832.1250]),
]
for x in inputs:
x = x.repeat(16, 16)
x = torch.log1p(x)
# Test argmax
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
assert metrics.generated_cpp_vec_kernel_count == 1
# Test argmin
torch._dynamo.reset()
metrics.reset()
self.common(fn2, (x,))
assert metrics.generated_cpp_vec_kernel_count == 1
# Currently, we enabled AVX2 and AVX512 for vectorization. If the platform is not
# supported, the vectorization will not work and skip this test case. For ARM or
# other platforms support, we just need to add the ISA info to the supported_vector_isa
# and include proper aten vectorization head file.
@unittest.skipIf(IS_FBCODE, "Not yet runnable in fbcode")
@requires_vectorization
@patch("torch.cuda.is_available", lambda: False)
def test_vec_kernel_cpu_only(self):
def fn(x1, x2):
# Current, there are some limitations as follows.
# rsqrt:
# assert [both a fallback and a decomp for same kernel: aten.rsqrt.default]
# round:
# couldn't find symbolic meta function/decomposition
# fmod/logical_and/logic_or:
# vec kernel has not support to_type
x = torch.abs(x1)
x = torch.sin(x)
x = torch.neg(x)
x = torch.square(x)
x = torch.sigmoid(x)
x = torch.relu(x)
x = torch.cos(x)
x = torch.exp(x)
x = torch.sqrt(x)
x = torch.add(x, x1)
x = torch.sub(x, x2)
x = torch.mul(x, x1)
x = torch.div(x, x1)
x = torch.pow(x, 10)
x = torch.log(x)
x = torch.floor(x)
x = torch.ceil(x)
x = torch.trunc(x)
x = torch.lgamma(x)
x = torch.fmod(x, x2)
x = torch.sign(x)
res = x + x2
return res
for dtype in vec_dtypes:
torch.manual_seed(0)
x1 = torch.randn((5, 20), dtype=dtype)
x2 = torch.randn((5, 20), dtype=dtype)
tol = 1e-2 if dtype == torch.bfloat16 else 1e-4
with config.patch({"cpp.simdlen": 1}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x1, x2))
assert metrics.generated_cpp_vec_kernel_count == 0
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x1, x2))
check_metrics_vec_kernel_count(1)
with config.patch({"cpp.simdlen": None}):
torch._dynamo.reset()
metrics.reset()
x1 = torch.randn(10, 20).permute(1, 0)
x2 = torch.randn((20, 10))
self.common(fn, (x1, x2))
check_metrics_vec_kernel_count(2)
torch._dynamo.reset()
metrics.reset()
x1 = torch.randn((10, 7))
x2 = torch.randn((10, 7))
self.common(fn, (x1, x2))
check_metrics_vec_kernel_count(1)
@unittest.skipIf(IS_FBCODE, "Not yet runnable in fbcode")
@unittest.skipIf(
sys.platform not in ["linux", "win32"],
"cpp kernel profile only support linux now",
)
@patch("torch.cuda.is_available", lambda: False)
@config.patch({"cpp.enable_kernel_profile": True})
@config.patch({"cpp.descriptive_names": "original_aten"})
def test_cpp_kernel_profile(self):
from torch.profiler import profile
@torch._dynamo.optimize("inductor", nopython=True)
def fn(a, b):
return a + b
a = torch.rand((100,))
b = torch.rand((100,))
with profile() as prof:
fn(a, b)
kernel_profile_events = []
for e in prof.profiler.function_events:
if "cpp_fused_add_0" in e.name:
kernel_profile_events.append(e.name)
assert len(kernel_profile_events) > 0
@requires_vectorization
def test_channel_shuffle_cl_output(self):
"""code and shape extracted from shufflenet_v2_x1_0"""
def channel_shuffle(x, groups):
batchsize, num_channels, height, width = x.size()
channels_per_group = num_channels // groups
x = x.view(batchsize, groups, channels_per_group, height, width)
x = torch.transpose(x, 1, 2).contiguous()
x = x.view(batchsize, -1, height, width)
return x.contiguous(memory_format=torch.channels_last)
for simdlen in (None, 256, 1):
with config.patch({"cpp.simdlen": simdlen}):
torch._dynamo.reset()
metrics.reset()
x = torch.randn(64, 58, 28, 28)
self.common(channel_shuffle, (x, 2))
if simdlen != 1:
check_metrics_vec_kernel_count(2)
@slowTest
@requires_vectorization
def test_transpose_with_norm(self):
"""a sub-module from TIMM gmlp_s16_224"""
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear = torch.nn.Linear(
in_features=256, out_features=1536, bias=True
)
self.act = torch.nn.GELU()
self.norm = torch.nn.LayerNorm(768)
self.proj = torch.nn.Linear(196, 196)
self.fc = torch.nn.Linear(in_features=768, out_features=256, bias=True)
def forward(self, x):
x = self.linear(x)
x = self.act(x)
u, v = x.chunk(2, dim=-1)
v = self.norm(v)
v = self.proj(v.transpose(-1, -2))
y = u * v.transpose(-1, -2)
return self.fc(y)
x = torch.randn(128, 196, 256)
for simdlen in (None, 256, 1):
with config.patch({"cpp.simdlen": simdlen}):
for eval_mode in [True, False]:
torch._dynamo.reset()
metrics.reset()
m = Model().eval() if eval_mode else Model()
self.common(m, (x,))
if simdlen != 1:
check_metrics_vec_kernel_count(8)
@requires_vectorization
@config.patch("cpp.enable_tiling_heuristics", False)
def test_transpose_copy(self):
def fn(a):
return a.t().contiguous()
for simdlen in (None, 256, 1):
with config.patch({"cpp.simdlen": simdlen}):
for dtype in (torch.float, torch.bfloat16):
for shape in (
(7, 7),
(8, 8),
(9, 9),
(16, 16),
(17, 17),
(32, 32),
(33, 33),
):
torch._dynamo.reset()
metrics.reset()
x = torch.randn(shape, dtype=dtype)
self.common(fn, (x,))
if simdlen != 1:
check_metrics_vec_kernel_count(2)
@torch._dynamo.config.patch(specialize_int=False)
def test_slice_scatter_issue122291(self):
@torch.compile(fullgraph=True)
def fn(t, t_src, dim, start, end, step):
return t.slice_scatter(t_src, dim, start, end, step)
shape = ((16, 16), (16, 2), 1, 4, 10, 1)
input_tensor = torch.zeros(shape[0], requires_grad=False, device="cpu")
src_tensor = torch.ones(shape[1], requires_grad=False, device="cpu")
with self.assertRaisesRegex(
torch._dynamo.exc.BackendCompilerFailed, r".*shape error in scatter op"
):
fn(input_tensor, src_tensor, shape[2], shape[3], shape[4], shape[5])
def test_horizontal_fusion(self):
def fn(a, b, c, idx):
_a = torch.index_select(a, dim=0, index=idx)
_b = torch.index_select(b, dim=0, index=idx)
_c = torch.index_select(c, dim=0, index=idx)
return _a, _b, _c
with config.patch({"cpp.max_horizontal_fusion_size": 0}):
metrics.reset()
torch._dynamo.reset()
a = torch.randn(size=(4, 16), dtype=torch.bfloat16)
b = torch.randn(size=(4, 16), dtype=torch.bfloat16)
c = torch.randn(size=(4, 16), dtype=torch.bfloat16)
idx = torch.zeros(size=[4], dtype=torch.int64)
opt_fn = torch._dynamo.optimize("inductor")(fn)
opt_fn(a, b, c, idx)
self.assertEqual(metrics.generated_kernel_count, 3)
self.assertTrue(same(fn(a, b, c, idx), opt_fn(a, b, c, idx)))
with config.patch({"cpp.max_horizontal_fusion_size": 1}):
metrics.reset()
torch._dynamo.reset()
a = torch.randn(size=(4, 32), dtype=torch.bfloat16)
b = torch.randn(size=(4, 32), dtype=torch.bfloat16)
c = torch.randn(size=(4, 32), dtype=torch.bfloat16)
idx = torch.zeros(size=[4], dtype=torch.int64)
opt_fn = torch._dynamo.optimize("inductor")(fn)
opt_fn(a, b, c, idx)
self.assertEqual(metrics.generated_kernel_count, 3)
self.assertTrue(same(fn(a, b, c, idx), opt_fn(a, b, c, idx)))
with config.patch({"cpp.max_horizontal_fusion_size": 2}):
metrics.reset()
torch._dynamo.reset()
a = torch.randn(size=(4, 64), dtype=torch.bfloat16)
b = torch.randn(size=(4, 64), dtype=torch.bfloat16)
c = torch.randn(size=(4, 64), dtype=torch.bfloat16)
idx = torch.zeros(size=[4], dtype=torch.int64)
opt_fn = torch._dynamo.optimize("inductor")(fn)
opt_fn(a, b, c, idx)
print(metrics.generated_kernel_count)
self.assertEqual(metrics.generated_kernel_count, 2)
self.assertTrue(same(fn(a, b, c, idx), opt_fn(a, b, c, idx)))
with config.patch({"cpp.max_horizontal_fusion_size": 3}):
metrics.reset()
torch._dynamo.reset()
a = torch.randn(size=(4, 128), dtype=torch.bfloat16)
b = torch.randn(size=(4, 128), dtype=torch.bfloat16)
c = torch.randn(size=(4, 128), dtype=torch.bfloat16)
idx = torch.zeros(size=[4], dtype=torch.int64)
opt_fn = torch._dynamo.optimize("inductor")(fn)
opt_fn(a, b, c, idx)
self.assertEqual(metrics.generated_kernel_count, 1)
self.assertTrue(same(fn(a, b, c, idx), opt_fn(a, b, c, idx)))
def test_lowp_fp_neg_abs(self):
def fn(x):
return x.neg().abs()
for dtype in _lowp_fp_dtypes:
metrics.reset()
x = torch.randn(100, 100).to(dtype)
opt_fn = torch._dynamo.optimize("inductor")(fn)
self.assertTrue(same(fn(x), opt_fn(x)))
assert metrics.cpp_to_dtype_count == 0
check_metrics_vec_kernel_count(1)
@config.patch("cpp.enable_tiling_heuristics", False)
def test_transpose_non_contiguous(self):
def fn(a):
# From part of timm HaloAttn:
# (https://github.com/rwightman/pytorch-image-models/blob/main/timm/layers/halo_attn.py#L97).
# Fixed https://github.com/pytorch/pytorch/issues/94269 accuracy issue.
as_strided = torch.ops.aten.as_strided.default(
a, [1, 384, 2, 20, 12], [153600, 1, 61440, 384, 7680]
)
as_strided_1 = torch.ops.aten.as_strided.default(
as_strided,
[1, 384, 2, 2, 12, 12],
[153600, 1, 61440, 3072, 7680, 384],
)
clone_1 = torch.ops.aten.clone.default(
as_strided_1, memory_format=torch.contiguous_format
)
_unsafe_view_1 = torch.ops.aten._unsafe_view.default(
clone_1, [8, 48, 4, 144]
)
permute_2 = torch.ops.aten.permute.default(_unsafe_view_1, [0, 2, 3, 1])
split_with_sizes = torch.ops.aten.split_with_sizes.default(
permute_2, [16, 32], -1
)
getitem = split_with_sizes[0]
getitem_1 = split_with_sizes[1]
permute_3 = torch.ops.aten.permute.default(getitem, [0, 1, 3, 2])
expand_1 = torch.ops.aten.expand.default(permute_3, [8, 4, 16, 144])
clone_3 = torch.ops.aten.clone.default(
expand_1, memory_format=torch.contiguous_format
)
return clone_3
metrics.reset()
x = torch.randn(1, 384, 20, 20).to(memory_format=torch.channels_last)
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_non_contiguous_index_with_constant_stride(self):
def fn(x):
x1 = x[:, :, :, ::2]
x2 = x[:, :, :, 1::2]
x = torch.stack((-x2, x1), dim=-1)
return x.flatten(-2)
metrics.reset()
x = torch.randn(1, 32, 16, 68)
opt_fn = torch._dynamo.optimize("inductor")(fn)
_, code = run_and_get_cpp_code(opt_fn, x)
self.assertTrue(same(fn(x), opt_fn(x)))
# def and use
FileCheck().check_count("cpp_fused", 2, exactly=True).run(code)
def test_invalid_index_of_empty_tensor(self):
def fn(a):
b = a[[0]]
return b
a = torch.tensor([])
with self.assertRaises(RuntimeError):
torch.compile(fn)(a)
@torch.no_grad()
@torch._inductor.config.patch(freezing=True)
def test_issue122380(self):
def func(x):
t1 = torch.unbind(x)
t2 = torch.stack(t1, dim=1)
t3 = torch.tanh(t2)
return t3
x = torch.randn(2, 3, 4)
self.assertEqual(torch.compile(func)(x), func(x))
@config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
def test_ir_node_str(self):
@torch.compile
def fn(x: torch.Tensor) -> torch.Tensor:
return x.sin(), torch.nn.Softmax(dim=1)(x.cos())
def run_node_alt(*args, **kwargs):
rv = run_node(*args, **kwargs)
strings.append(str(rv))
return rv
strings = []
run_node = GraphLowering.run_node
with patch.object(GraphLowering, "run_node", run_node_alt):
fn(torch.randn([8, 128]))
self.assertGreater(len(strings), 3)
def test_vertical_sum_cpu_only(self):
def fn1(a):
return a.sum(dim=0)
def fn2(a):
return a.sum(dim=1)
metrics.reset()
x = torch.randn(100, 100)
self.common(fn1, (x,))
check_metrics_vec_kernel_count(1)
metrics.reset()
x = torch.randn(100, 100, 100)
self.common(fn2, (x,))
check_metrics_vec_kernel_count(1)
def test_transpose_vertical_sum_cpu_only(self):
def fn(a, b):
c = a * b
return c.sum(dim=1)
metrics.reset()
x = torch.randn(100, 50, 50)
y = torch.randn(100, 50, 50).transpose(1, 2)
self.common(fn, (x, y))
check_metrics_vec_kernel_count(2)
def test_transpose_mxn_16_16_bf16_fp16(self):
def fn(a, b):
c = a * b
return c.sum(dim=1)
for dtype in [torch.bfloat16, torch.float16]:
metrics.reset()
x = torch.randn(100, 50, 50).to(dtype)
y = torch.randn(100, 50, 50).to(dtype).transpose(1, 2)
self.common(fn, (x, y))
check_metrics_vec_kernel_count(2)
def test_transpose_mxn_32_32_bf16_fp16(self):
def fn(a):
return a.permute(0, 2, 1).contiguous()
for dtype in [torch.bfloat16, torch.float16]:
metrics.reset()
x = torch.randn(2, 9216, 9216).to(dtype)
self.common(fn, (x,))
check_metrics_vec_kernel_count(2)
def test_transpose_sum2d_cpu_only(self):
def fn(a, b):
c = a * b
return c.sum()
metrics.reset()
x = torch.randn(50, 50)
y = torch.randn(50, 50).transpose(0, 1)
self.common(fn, (x, y))
check_metrics_vec_kernel_count(2)
@config.patch("cpp.enable_tiling_heuristics", False)
def test_transpose_sum_outer(self):
# https://github.com/pytorch/pytorch/issues/98573
def fn(a):
return a.transpose(2, 3).sum(dim=1).contiguous()
metrics.reset()
x = torch.randn(10, 50, 50, 50)
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_to_dtype_bool_float(self):
# https://github.com/pytorch/pytorch/issues/100800
def f(a):
return torch.where(
torch.ones_like(a).to(torch.bool),
torch.zeros_like(a),
torch.ones_like(a) * 2,
)
self.common(f, (torch.ones(16),))
def test_to_dtype_float_bool(self):
# https://github.com/pytorch/pytorch/issues/100466
def f(a):
a = a * torch.tensor(a >= 0, dtype=torch.float32)
return a
x = torch.rand(16)
self.common(f, (x,))
def test_constant_store(self):
# https://github.com/pytorch/pytorch/issues/104515
def f(a):
a[0, [3, 3]] = -float("inf")
return a
x = torch.rand(4, 5)
self.common(f, (x,))
def test_to_channels_last_lowp_fp(self):
def f(a):
return a.to(memory_format=torch.channels_last)
for dtype in _lowp_fp_dtypes:
x = torch.rand(2, 3, 14, 14).to(dtype)
self.common(f, (x,))
def test_broadcast_mul_lowp_fp(self):
def f(a, b):
return a * b
for dtype in _lowp_fp_dtypes:
a = torch.randn(2, 16, 16).to(dtype)
b = torch.randn(2, 1, 1).to(dtype)
self.common(f, (a, b))
def test_linear_buffer_reuse(self):
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear1 = torch.nn.Linear(16, 16)
self.tanh = torch.nn.Tanh()
self.linear2 = torch.nn.Linear(16, 16)
def forward(self, x):
x = self.linear1(x)
x = self.tanh(x)
x = self.linear2(x)
return x
mod = M().eval()
v = torch.randn(1, 16)
with torch.no_grad():
def compile_fx_wrapper(model_, example_inputs_):
return compile_fx(model_, example_inputs_)
def run(*ex, **kwargs):
return mod(*ex, **kwargs)
run = torch._dynamo.optimize(compile_fx_wrapper)(run)
_, code = run_and_get_cpp_code(run, v)
self.assertFalse("= as_strided(" in code)
self.assertEqual(run(*v), mod(*v))
def test_invalid_dropout_args(self):
class MyModel(torch.nn.Module):
def forward(self, x):
x = x * 2
x = torch.nn.functional.dropout(x, p=0.5)
x = torch.relu(x)
return x
example_inputs = torch.tensor([[1, 2, 3], [4, 5, 6]])
func = MyModel()
jit_func = torch.compile(func)
self.assertRaises(RuntimeError, lambda: func(example_inputs))
self.assertRaises(RuntimeError, lambda: jit_func(example_inputs))
def test_nn_param_assign(self):
# https://github.com/pytorch/pytorch/issues/99569
class Model2(nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv = nn.Conv2d(in_channels=3, out_channels=5, kernel_size=3)
self.batchnorm = nn.BatchNorm2d(num_features=5)
self.conv_weight = torch.randn(5, 3, 3, 3)
self.conv_bias = torch.randn(5)
def forward(self, x):
self.conv.weight = nn.Parameter(self.conv_weight)
self.conv.bias = nn.Parameter(self.conv_bias, requires_grad=False)
self.conv.eval()
x = self.conv(x)
x = self.batchnorm(x)
x = F.relu(x)
return x
input_tensor = torch.randn(1, 3, 10, 10)
func = Model2().to("cpu")
with torch.no_grad():
func.train(False)
v1 = func(input_tensor)
jit_func = torch.compile(func, fullgraph=True)
v2 = jit_func(input_tensor)
self.assertEqual(v1, v2)
def test_nn_param_assign_wrapped(self):
class Model2(nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv = nn.Conv2d(in_channels=3, out_channels=5, kernel_size=3)
self.batchnorm = nn.BatchNorm2d(num_features=5)
self.conv_weight = torch.randn(5, 3, 3, 3)
self.conv_bias = torch.randn(5)
def forward(self, x):
self.conv.weight = nn.Parameter(self.conv_weight)
self.conv.bias = nn.Parameter(self.conv_bias, requires_grad=False)
self.conv.eval()
x = self.conv(x)
x = self.batchnorm(x)
x = F.relu(x)
return x
input_tensor = torch.randn(1, 3, 10, 10)
func = Model2().to("cpu")
@functools.wraps(func)
def wrapper(*args, **kwargs):
return func(*args, **kwargs)
with torch.no_grad():
func.train(False)
v1 = func(input_tensor)
jit_func = torch.compile(wrapper, fullgraph=True)
v2 = jit_func(input_tensor)
self.assertEqual(v1, v2)
@config.patch(inplace_buffers=True)
def test_in_out_buffer(self):
def fn(x, y):
z = torch.matmul(x, y.transpose(-1, -2)) / 8.0
return z
inps = [torch.randn(1, 2, 8, 4), torch.randn(1, 2, 8, 4)]
fn_opt = torch._dynamo.optimize("inductor")(fn)
_, code = run_and_get_cpp_code(fn_opt, *inps)
self.assertTrue("in_out_ptr" in code)
self.assertEqual(fn_opt(*inps), fn(*inps))
def test_eliminate_meaningless_copy(self):
def fn(x1, x2):
permute = torch.ops.aten.permute.default(x2, [0, 2, 1, 3])
clone = torch.ops.aten.clone.default(
permute, memory_format=torch.contiguous_format
)
view = torch.ops.aten.view.default(clone, [1024, -1, 32])
bmm = torch.ops.aten.bmm.default(view, x1)
permute = torch.ops.aten.permute.default(view, [0, 2, 1])
return (bmm, permute)
metrics.reset()
self.common(
fn,
[
rand_strided(
(1024, 32, 128), (4096, 1, 32), device="cpu", dtype=torch.float32
),
rand_strided(
(64, 128, 16, 32),
(65536, 512, 32, 1),
device="cpu",
dtype=torch.float32,
),
],
)
self.assertEqual(metrics.generated_kernel_count, 1)
def test_attention_size_mismatch(self):
class Attention(torch.nn.Module):
def __init__(self, hidden_size, num_heads):
super().__init__()
self.hidden_size = hidden_size
self.num_heads = num_heads
self.head_size = hidden_size // num_heads
self.query = torch.nn.Linear(hidden_size, hidden_size)
self.key = torch.nn.Linear(hidden_size, hidden_size)
self.value = torch.nn.Linear(hidden_size, hidden_size)
self.inv_scale = torch.nn.Parameter(
torch.Tensor([1 / self.head_size**0.5]), requires_grad=False
)
def forward(self, x):
query = self.query(x)
key = self.key(x)
value = self.value(x)
(batch_size, seq_len, hidden_size) = query.size()
query = query.view(
batch_size, seq_len, self.num_heads, self.head_size
).permute(0, 2, 1, 3)
key = key.view(
batch_size, seq_len, self.num_heads, self.head_size
).permute(0, 2, 3, 1)
value = value.view(
batch_size, seq_len, self.num_heads, self.head_size
).permute(0, 2, 1, 3)
attention_weights = (
torch.matmul(query, key).mul(self.inv_scale).softmax(dim=-1)
)
output = torch.matmul(attention_weights, value)
return output
torch.manual_seed(123)
hidden_size = 16
num_heads = 1
seq_len = 4
batch_size = 1
x = torch.randn(batch_size, seq_len, hidden_size)
func = Attention(hidden_size, num_heads).to("cpu")
with torch.no_grad():
res1 = func(x)
jit_func = torch.compile(func)
res2 = jit_func(x)
self.assertEqual(res1, res2)
def test_scalar_mul_bfloat16(self):
def f(x):
return torch.ops.aten.mul.Tensor(x, 1.7015043497085571)
metrics.reset()
x = torch.randn(4, 5, dtype=torch.bfloat16)
self.common(f, (x,))
check_metrics_vec_kernel_count(1)
def test_bf16_zeros(self):
def fn():
x = torch.zeros(1, 1, 32, dtype=torch.bfloat16)
return x
self.common(fn, ())
def test_select_tiliing_with_index_expr(self):
def fn(x, y):
x = torch.ops.aten.view.default(x, [8, 8, 8, 3136])
x = torch.ops.aten.permute.default(x, [0, 1, 3, 2])
y = torch.ops.aten.mul.Tensor(y, x)
return torch.ops.aten.constant_pad_nd.default(y, [0, 0, 1, 0, 0, 0], 0.0)
x = torch.randn(8, 64, 56, 56)
y = torch.randn(8, 8, 3136, 8)
self.common(fn, (x, y))
@unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
@patch("torch.cuda.is_available", lambda: False)
@config.patch(freezing=True)
def test_linear_with_no_default_contiguous_input(self):
dtypes = [
torch.float32,
]
if torch.ops.mkldnn._is_mkldnn_bf16_supported():
dtypes.append(torch.bfloat16)
if torch.ops.mkldnn._is_mkldnn_fp16_supported():
dtypes.append(torch.float16)
mod = torch.nn.Sequential(torch.nn.Linear(16, 16)).eval()
temp = torch.randn(1, 16, 1, 1)
v = torch.as_strided(temp, [1, 16], [0, 1], 0)
self.assertTrue(v.is_contiguous())
for dtype in dtypes:
with torch.no_grad():
self.common(
mod.to(dtype),
(v.to(dtype),),
)
@patch("torch.cuda.is_available", lambda: False)
@config.patch(freezing=True)
def test_linear_with_reshape(self):
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear = torch.nn.Linear(16, 16, bias=False)
def forward(self, x):
x = self.linear(x)
return x.view(4, 4, 4)
mod = M().eval()
v = torch.randn(4, 16)
with torch.no_grad():
torch._dynamo.reset()
metrics.reset()
self.common(
mod,
(v,),
)
assert metrics.generated_kernel_count == 0
@config.patch(implicit_fallbacks=True)
def test_aten_normal_dtype(self):
for dtype in [torch.float64, torch.float16, None]:
def fn():
return torch.normal(2, 3, (10, 10), dtype=dtype, device="cpu")
self.assertEqual(
torch.compile(fn, backend="aot_eager_decomp_partition")().dtype,
dtype if dtype else torch.float32,
)
self.assertEqual(
torch.compile(fn, backend="inductor")().dtype,
dtype if dtype else torch.float32,
)
def test_group_norm_vec(self):
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.group_norm = torch.nn.GroupNorm(3, 90)
def forward(self, x):
return self.group_norm(x)
options = itertools.product(
vec_dtypes, [torch.contiguous_format, torch.channels_last]
)
for dtype, fmt in options:
torch._dynamo.reset()
metrics.reset()
mod = M().eval()
x = torch.randn((2, 90, 6, 6), dtype=dtype).to(memory_format=fmt)
with torch.no_grad():
self.common(mod, (x,))
# 2 generated kernels (one for var_mean, the other for result)
check_metrics_vec_kernel_count(2)
# check loop split optimization
if fmt == torch.channels_last:
torch._dynamo.reset()
metrics.reset()
with torch.no_grad():
opt_mod = torch.compile(mod)
_, code = run_and_get_cpp_code(opt_mod, x)
# check that there are no non_contiguous loads
FileCheck().check_count("__at_align__ std::array", 0, exactly=True).run(
code
)
def test_int_div_vec(self):
def fn(x, y, mode):
return torch.div(x, y, rounding_mode=mode)
for dtype in [
torch.int8,
torch.uint8,
torch.int32,
torch.int64,
]:
x = torch.randint(1, 100, (32, 32), dtype=dtype)
y = torch.randint(1, 100, (32, 32), dtype=dtype)
for mode in [None, "trunc", "floor"]:
with torch.no_grad():
metrics.reset()
self.common(fn, (x, y, mode))
check_metrics_vec_kernel_count(1)
def test_uint8_add(self):
# https://github.com/pytorch/pytorch/issues/113016
def fn(x, y):
return torch.add(x, y).neg().to(torch.int32)
x = torch.randint(0, 255, (3, 3), dtype=torch.uint8)
y = torch.randint(0, 255, (3, 3), dtype=torch.uint8)
self.common(fn, (x, y))
def test_uint8_sub(self):
# https://github.com/pytorch/pytorch/issues/113016
def fn(x, y):
return torch.sub(x, y).neg().to(torch.int32)
x = torch.randint(0, 255, (3, 3), dtype=torch.uint8)
y = torch.randint(0, 255, (3, 3), dtype=torch.uint8)
self.common(fn, (x, y))
def test_non_contiguous_reduction_store(self):
# https://github.com/pytorch/pytorch/issues/113018
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv = torch.nn.Conv2d(39, 1, kernel_size=(1, 17), stride=(2, 2))
def forward(self, x):
return self.conv(x.max(3).values)
m = M()
x = torch.randn(1, 39, 1, 18, 17)
self.common(m, (x,))
def test_embedding_vec(self):
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.emb = torch.nn.Embedding(64, 128)
def forward(self, idx, x):
return self.emb(idx) + x
idx = torch.randint(0, 64, (4, 32))
x = torch.randn(4, 32, 128)
m = M().eval()
with torch.no_grad():
metrics.reset()
self.common(m, (idx, x))
check_metrics_vec_kernel_count(1)
def test_embedding_vec_bf16(self):
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.emb = torch.nn.Embedding(64, 128)
def forward(self, idx, x):
return self.emb(idx)
idx = torch.randint(0, 64, (4, 32))
x = torch.randn(4, 32, 128).to(torch.bfloat16)
m = M().eval()
with torch.no_grad():
metrics.reset()
self.common(m, (idx, x))
check_metrics_vec_kernel_count(1)
# we are doing direct load/store, make sure we do not generate
# redundant type casts
m_opt = torch.compile(m)
_, code = run_and_get_cpp_code(m_opt, idx, x)
self.assertTrue("Vectorized" in code)
self.assertTrue("cvt_lowp_fp_to_fp32" not in code)
self.assertTrue("cvt_fp32_to_lowp_fp" not in code)
def test_concat_inner_vec(self):
def fn(x, y):
return F.relu(torch.cat([x, y], dim=1))
x = torch.randn(32, 35)
y = torch.randn(32, 120)
metrics.reset()
self.common(fn, (x, y))
check_metrics_vec_kernel_count(3)
@config.patch("cpp.enable_tiling_heuristics", False)
def test_expr_vec_non_contiguous(self):
def fn(x):
# the pattern from sebotnet33ts_256
y = torch.nn.functional.pad(x, (0, 31)).reshape(-1, 33, 63)
y = y[:, :32, 31:].reshape(4, 32, 1, 32, 32).expand(-1, -1, 32, -1, -1)
y = y.permute(0, 3, 1, 4, 2).clone(memory_format=torch.contiguous_format)
y = y.view(4, 1024, 1024)
return y.softmax(dim=-1)
x = torch.randn(128, 2048)
opt_fn = torch.compile(fn)
metrics.reset()
_, code = run_and_get_cpp_code(opt_fn, x)
self.assertTrue(same(fn(x), opt_fn(x)))
# 4 kernels for max, exp, sum and div
check_metrics_vec_kernel_count(4)
FileCheck().check_count(
"Vectorized<int>::loadu(tmpbuf.data())", 0, exactly=True
).run(code)
def test_vec_contiguous_ModularIndexing(self):
# https://github.com/pytorch/pytorch/issues/114488
class M(torch.nn.Module):
def __init__(self, dim):
super().__init__()
self.norm = torch.nn.LayerNorm(dim * 4)
def forward(self, x):
# the pattern from swin_base_patch4_window7_224
B, H, W, C = x.shape
x = (
x.reshape(B, H // 2, 2, W // 2, 2, C)
.permute(0, 1, 3, 4, 2, 5)
.flatten(3)
)
x = self.norm(x)
return x
x = torch.randn(1, 56, 56, 128)
m = M(128)
opt_m = torch.compile(m)
with torch.no_grad():
metrics.reset()
_, code = run_and_get_cpp_code(opt_m, x)
self.assertTrue(same(m(x), opt_m(x)))
# Two kernels: one for reduction, one pointwises
check_metrics_vec_kernel_count(2)
FileCheck().check_count(
"Vectorized<float>::loadu(tmpbuf.data())", 0, exactly=True
).run(code)
@parametrize("dtype", (torch.float16, torch.bfloat16, torch.float))
@parametrize("shape", ("15,3,13", "4,2048,4096"))
def test_fp8_cast(self, dtype: torch.dtype, shape: str):
def fp8_cast(x):
y0 = x.to(dtype=torch.float8_e4m3fn).to(dtype)
y1 = x.to(dtype=torch.float8_e5m2).to(dtype)
return y0, y1
shape = [int(dim) for dim in shape.split(",")]
x = torch.rand(*shape, device="cpu", dtype=dtype)
self.common(fp8_cast, (x,))
def test_logical_op_store_to_lowp_data_dtype(self):
# https://github.com/pytorch/pytorch/issues/117624
# https://github.com/pytorch/pytorch/issues/117627
def fn(out1, out2, input, other):
o1 = torch.logical_or(out=out1, input=input, other=other)
o2 = torch.logical_xor(out=out2, input=input, other=other)
return o1, o2
x = torch.rand([3, 3, 2, 8, 9, 2], dtype=torch.float)
y = torch.rand([3, 3, 2, 8, 9, 2], dtype=torch.float)
for dtype in _lowp_fp_dtypes:
o1 = torch.rand([3, 3, 2, 8, 9, 2], dtype=dtype)
o2 = torch.rand([3, 3, 2, 8, 9, 2], dtype=dtype)
with torch.no_grad():
self.common(fn, (o1, o2, x, y))
def test_constant_bool_vec(self):
def fn(x):
mask = torch.zeros(1, dtype=torch.bool)
return torch.where(mask, x, -1.0)
x = torch.rand(1000)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
@torch._dynamo.config.patch(dynamic_shapes=True)
@torch._dynamo.config.patch(assume_static_by_default=False)
def test_symbolic_shape_scalar_value_reduction(self):
def fn(x, y):
return y + torch.ones(x).sum()
with torch.no_grad():
metrics.reset()
y = torch.randn(100)
self.common(fn, (100, y))
check_metrics_vec_kernel_count(2)
def test_int32_pointwise_vec(self):
def fn(x):
return x * x
x = torch.randint(0, 100, (32, 32), dtype=torch.int32)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_int32_reduction_vec(self):
def fn(x):
return x.sum(dim=1)
x = torch.randint(0, 100, (32, 32), dtype=torch.int32)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_uint32_pointwise_vec(self):
def fn(x):
return x * x
x = torch.randint(0, 100, (32, 32), dtype=torch.uint32)
metrics.reset()
self.common(fn, (x,))
# TODO(jgong5): change to 1 with vectorized uint32 load
assert metrics.generated_cpp_vec_kernel_count == 0
def test_uint32_reduction_vec(self):
def fn(x):
return x.sum(dim=1)
x = torch.randint(0, 100, (32, 32), dtype=torch.uint32)
metrics.reset()
self.common(fn, (x,))
# TODO(jgong5): change to 1 with vectorized uint32/uint64 load
assert metrics.generated_cpp_vec_kernel_count == 0
def test_int64_pointwise_vec(self):
def fn(x):
return x * x
x = torch.randint(0, 100, (32, 32), dtype=torch.int64)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_int64_reduction_vec(self):
def fn(x):
return x.sum(dim=1)
x = torch.randint(0, 100, (32, 32), dtype=torch.int64)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_uint64_pointwise_vec(self):
def fn(x):
return x * x
x = torch.randint(0, 100, (32, 32), dtype=torch.uint64)
metrics.reset()
self.common(fn, (x,))
# TODO(jgong5): change to 1 with vectorized uint64 load
assert metrics.generated_cpp_vec_kernel_count == 0
def test_uint64_reduction_vec(self):
def fn(x):
return x.sum(dim=1)
x = torch.randint(0, 100, (32, 32), dtype=torch.uint64)
metrics.reset()
self.common(fn, (x,))
# TODO(jgong5): change to 1 with vectorized uint64 load
assert metrics.generated_cpp_vec_kernel_count == 0
def test_convert_int32_to_int64_vec(self):
def fn(x):
return x.to(torch.int64)
x = torch.randint(0, 100, (32, 32), dtype=torch.int32)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_convert_int64_to_int32_vec(self):
def fn(x):
return x.to(torch.int32)
x = torch.randint(0, 100, (32, 32), dtype=torch.int64)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_convert_fp32_to_int64_vec(self):
def fn(x):
return x.to(torch.int64)
x = torch.rand(32, 32)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_convert_int64_to_fp32_vec(self):
def fn(x):
return x.to(torch.float32)
x = torch.randint(0, 100, (32, 32), dtype=torch.int64)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_double_pointwise_vec(self):
def fn(x):
return x * x
x = torch.randn((32, 32), dtype=torch.double)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_double_reduction_vec(self):
def fn(x):
return x.sum(dim=1)
x = torch.randn((32, 32), dtype=torch.double)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_convert_fp32_to_double_vec(self):
def fn(x):
return x.to(torch.double)
x = torch.randn(32, 32)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_convert_double_to_fp32_vec(self):
def fn(x):
return x.to(torch.float32)
x = torch.randn((32, 32), dtype=torch.double)
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
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)
class Model1(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
def forward(self, *args):
cat = torch.cat((args[3], args[2], args[1], args[0]), dim=2)
max_1 = torch.max(args[4], p0)
mul = torch.mul(cat, max_1)
tan = torch.tan(mul)
return (mul, tan)
metrics.reset()
m = Model1()
self.common(
m,
(
torch.randn((17, 5, 1, 7)).half(),
torch.randn((17, 5, 1, 7)).half(),
torch.randn((17, 5, 11, 7)).half(),
torch.randn((17, 5, 1, 7)).half(),
torch.tensor(4.39, dtype=torch.float16),
),
)
def test_masked_load_int64_vec(self):
# https://github.com/pytorch/pytorch/issues/120377
def fn(x):
return torch.nn.functional.pad(x, (0, 13))
x = torch.randint(0, 100, (819,), dtype=torch.int64)
metrics.reset()
self.common(fn, (x,))
assert metrics.generated_cpp_vec_kernel_count == 1
def test_highp_to_lowp_cse_var_cache_with_store(self):
# Fix issue: https://github.com/pytorch/pytorch/issues/128263
input = torch.randn(5, 128, dtype=torch.float32)
input2 = torch.randint(0, 10, (5, 128), dtype=torch.int8)
input3 = torch.randn(128, 128, dtype=torch.float32)
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
def forward(self, x, x2, x3):
x2 = x2.to(torch.int32)
temp = test_operators.realize(x2.to(torch.float16))
temp2 = temp.to(torch.float32)
temp2 = temp2 * x
return torch.mm(temp, x3.to(torch.float16)), temp2
metrics.reset()
m = Model()
self.common(
m,
(input, input2, input3),
)
def test_reduction_float_to_int64(self):
# https://github.com/pytorch/pytorch/issues/124821
def fn(x):
return x.max(0).values
x = torch.randint(0, 100, (22, 51), dtype=torch.int64)
metrics.reset()
self.common(fn, (x,))
assert metrics.generated_cpp_vec_kernel_count == 1
@config.patch({"cpp.dynamic_threads": True})
def test_reduction_with_dynamic_threads(self):
def fn(a, b):
return a.sum(), b.sum()
self.common(
fn,
(torch.randn(1000), torch.rand(1000)),
)
@patch("torch.cuda.is_available", lambda: False)
@config.patch(freezing=True)
def test_linear_float64(self):
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.weight1 = torch.nn.Parameter(
torch.randn(10, 10, dtype=torch.float64)
)
self.weight2 = torch.nn.Parameter(
torch.randn(10, 10, dtype=torch.float64)
)
self.bias = torch.nn.Parameter(torch.randn(10, dtype=torch.float64))
def forward(self, x1):
v1 = torch.mm(x1, self.weight1)
v2 = torch.addmm(self.bias, x1, self.weight2)
return (v1, v2)
mod = M().eval()
v = torch.randn(10, 10, dtype=torch.float64)
with torch.no_grad():
self.common(
mod,
(v,),
)
def test_fused_attention_conv(self):
# https://github.com/pytorch/pytorch/issues/121174.
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.q_conv = torch.nn.Conv2d(4, 4, 1)
self.k_conv = torch.nn.Conv2d(4, 4, 1)
self.v_conv = torch.nn.Conv2d(4, 4, 1)
def forward(self, x):
q = self.q_conv(x)
k = self.k_conv(x)
v = self.v_conv(x)
q = q.permute(0, 2, 1, 3)
k = k.permute(0, 2, 1, 3)
v = v.permute(0, 2, 1, 3)
return torch.nn.functional.scaled_dot_product_attention(
q, k, v, dropout_p=0.0, is_causal=False
)
fn = Model()
x = torch.randn(1, 4, 2, 2)
self.common(fn, (x,))
@parametrize("is_inference", (True, False))
def test_disabled_amp(self, is_inference):
class M(torch.nn.Module):
def __init__(self):
super().__init__()
self.all_head_size = 12 * 64
self.dense = nn.Linear(self.all_head_size, self.all_head_size)
def forward(self, q, k, v):
context_layer = F.scaled_dot_product_attention(
q, k, v, attn_mask=None, dropout_p=0.2
)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (
self.all_head_size,
)
context_layer = context_layer.view(new_context_layer_shape)
return self.dense(context_layer)
mod = M().to(torch.bfloat16).eval()
q = torch.randn((4, 12, 512, 64), dtype=torch.bfloat16) / 10.0
k = torch.randn((4, 12, 512, 64), dtype=torch.bfloat16) / 10.0
v = torch.randn((4, 12, 512, 64), dtype=torch.bfloat16) / 10.0
inputs = (
q,
k,
v,
)
compiler_mode = torch.compile(mod)
from torch.nn.attention import sdpa_kernel, SDPBackend
context = contextlib.nullcontext if not is_inference else torch.no_grad
with config.patch(
{"fallback_random": True}
), torch.cpu.amp.autocast(), context(), sdpa_kernel(SDPBackend.MATH):
torch.manual_seed(0)
eager = mod(*inputs)
torch.manual_seed(0)
self.assertEqual(compiler_mode(*inputs), eager)
@requires_vectorization
def test_vec_indirect_load_cse_cache(self):
# https://github.com/pytorch/pytorch/issues/123502
from math import inf
def fn(arg0_1):
full_default = torch.ops.aten.full.default([209985], 1)
select = torch.ops.aten.select.int(arg0_1, 0, 0)
select_1 = torch.ops.aten.select.int(arg0_1, 0, 1)
view = torch.ops.aten.reshape.default(select_1, [-1])
expand = torch.ops.aten.expand.default(view, [209985])
full_default_1 = torch.ops.aten.full.default([10000], 0)
scatter_add = torch.ops.aten.scatter_add.default(
full_default_1, 0, expand, full_default
)
pow_1 = torch.ops.aten.pow.Tensor_Scalar(scatter_add, -0.5)
eq = torch.ops.aten.eq.Scalar(pow_1, inf)
full_default_2 = torch.ops.aten.full.default([], 0.0)
where = torch.ops.aten.where.self(eq, full_default_2, pow_1)
index = torch.ops.aten.index.Tensor(where, [select])
index_1 = torch.ops.aten.index.Tensor(where, [select_1])
mul_1 = torch.ops.aten.mul.Tensor(index, index_1)
return (mul_1,)
x = torch.zeros(2, 209985).to(torch.int64)
opt_fn = torch._dynamo.optimize("inductor")(fn)
_, code = run_and_get_cpp_code(opt_fn, x)
FileCheck().check_count(
"return at::vec::VectorizedN<int64_t,2>::loadu(tmpbuf.data(),",
4,
exactly=True,
).run(code)
def test_load_half(self):
def fn(arg0_1, arg0_2):
return arg0_1.copy_(arg0_2)
with config.patch({"cpp.simdlen": 0}):
x1 = torch.randn(2, 10).to(torch.half)
x2 = torch.randn(2, 10).to(torch.half)
opt_fn = torch._dynamo.optimize("inductor")(fn)
_, code = run_and_get_cpp_code(opt_fn, x1, x2)
FileCheck().check_count(
"static_cast<float>",
0,
exactly=True,
).run(code)
def test_repeated_exp(self):
def fn(x):
y = x.sigmoid()
return y + 1, y.sum(-1)
x = torch.randn(1000, 1000)
opt_fn = torch.compile(fn)
_, code = run_and_get_cpp_code(opt_fn, x)
FileCheck().check_count(
".exp()",
1,
exactly=True,
).run(code)
def test_convert_fp32_int64_oob_vec(self):
# https://github.com/pytorch/pytorch/issues/129863
def fn(x):
float32 = x.to(torch.float32)
return float32.to(torch.int64)
x = torch.full((32,), -9223372036854775808, dtype=torch.int64)
for simdlen in (None, 256):
with config.patch({"cpp.simdlen": simdlen}):
torch._dynamo.reset()
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
def test_consistent_remove_buffers(self):
def fn(x):
z = x + x
z1 = test_operators.realize(z)
return x + z1
# The shape makes sure we generate both vec and scalar kernels
x = torch.randn((65,), dtype=torch.bfloat16)
with config.patch(inplace_buffers=False):
metrics.reset()
self.common(fn, (x,))
check_metrics_vec_kernel_count(1)
_, code = run_and_get_cpp_code(torch.compile(fn), x)
FileCheck().check_count(
"tmp1 + tmp2",
2,
exactly=True,
).run(code)
@requires_vectorization
def test_bool_reduction_vec(self):
for op in (
torch.masked.mean,
torch.any,
torch.min,
torch.max,
):
def fn(x1, x2, x3):
return op(x1), op(x2), op(x3)
c = [False] * 63
input1 = torch.Tensor(c).to(torch.bool)
c[10] = True
input2 = torch.Tensor(c).to(torch.bool)
input3 = torch.Tensor([True] * 63).to(torch.bool)
metrics.reset()
self.common(
fn,
(
input1,
input2,
input3,
),
)
n_veckernel = 6 if op is torch.masked.mean else 3
check_metrics_vec_kernel_count(n_veckernel)
@requires_vectorization
def test_full_bits_lowp(self):
def check_use_full_bits(func, shapes, dtype, mixed, check_vecn):
example_inputs = [torch.randn(shape, dtype=dtype) for shape in shapes]
if mixed:
example_inputs[0] = example_inputs[0].to(
dtype=torch.half if dtype == torch.bfloat16 else torch.bfloat16
)
f_opt = torch.compile()(func)
_, code = run_and_get_cpp_code(f_opt, *example_inputs)
if check_vecn:
self.assertTrue(
"at::vec::VectorizedN" in code or "at::vec::convert<float,2" in code
)
else:
self.assertFalse(
"at::vec::VectorizedN" in code or "at::vec::convert<float,2" in code
)
funcs = []
def func0(arg0, arg1):
return torch.ops.aten.sum(
torch.ops.aten.add(torch.ops.aten.atanh(arg0), arg1), (2, 3)
)
funcs.append(func0)
def func1(arg0):
v = torch.ops.prims.convert_element_type.default(arg0, torch.float)
v = torch.ops.aten.add(torch.ops.aten.atanh(arg0), v)
return torch.ops.prims.convert_element_type.default(v, arg0.dtype)
funcs.append(func1)
def func2(arg0, arg1):
v = torch.ops.aten.atanh(arg0)
v = torch.ops.aten.add(v, arg1)
return torch.ops.prims.convert_element_type.default(v, arg1.dtype)
funcs.append(func2)
# test small shapes
funcs.append(func2)
small_size = cpu_vec_isa.pick_vec_isa().nelements(dtype=torch.bfloat16) // 2
example_shapes = [
[(10, 32, 20, 20), (10, 32, 20, 20)],
[(10, 32, 20, 20)],
[(10, 32, 20, 20), (10, 32, 20, 20)],
# test small shapes
[(small_size), (small_size)],
]
mixed_types = [False, False, True, False]
check_vecns = [True, True, True, False]
for dtype in [torch.bfloat16, torch.float16]:
for func, shapes, mixed, check_vecn in zip(
funcs, example_shapes, mixed_types, check_vecns
):
check_use_full_bits(func, shapes, dtype, mixed, check_vecn)
if __name__ == "__main__":
from torch._inductor.test_case import run_tests
from torch.testing._internal.inductor_utils import HAS_CPU
if HAS_CPU and not IS_MACOS:
run_tests(needs="filelock")