Skip to content

tile dynamic dim #3085

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Aug 29, 2024
Merged

tile dynamic dim #3085

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -953,7 +953,7 @@ def aten_ops_cumsum(
)


@dynamo_tensorrt_converter(torch.ops.aten.tile.default)
@dynamo_tensorrt_converter(torch.ops.aten.tile.default, supports_dynamic_shapes=True)
@enforce_tensor_types(
{
0: (TRTTensor,),
Expand Down
63 changes: 59 additions & 4 deletions py/torch_tensorrt/dynamo/conversion/impl/slice/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -457,6 +457,7 @@ def tile(
dims: Sequence[int],
) -> TRTTensor:
diff = len(dims) - len(input.shape)
has_dynamic_shape_input = has_dynamic_shape(input.shape)
if diff > 0:
# prepend 1 to input.shape
new_shape = (1,) * diff + tuple(input.shape)
Expand All @@ -467,10 +468,64 @@ def tile(
# prepend 1 to dims
dims = (1,) * -diff + tuple(dims)

shapes = [i * j for i, j in zip(input.shape, dims)]
starts = [0] * len(dims)
strides = [1] * len(dims)
layer = ctx.net.add_slice(input, tuple(starts), tuple(shapes), tuple(strides))
starts = tuple([0] * len(dims))
strides = tuple([1] * len(dims))
# layer = ctx.net.add_slice(input, tuple(starts), tuple(shapes), tuple(strides))
if not (has_dynamic_shape_input):
shapes = [i * j for i, j in zip(input.shape, dims)]
layer = ctx.net.add_slice(input, tuple(starts), tuple(shapes), tuple(strides))
else:
shapes = []
index = 0
for i, j in zip(input.shape, dims):
if i == DYNAMIC_DIM:
i = get_shape(
ctx, target, source_ir, name + f"_input_{index}", input, index
)
prod_shape = convert_binary_elementwise(
ctx,
target,
source_ir,
name + "_prod",
trt.ElementWiseOperation.PROD,
i,
j,
)
shapes.append(prod_shape)
index = index + 1
layer = ctx.net.add_slice(
input, start=trt.Dims(), shape=trt.Dims(), stride=trt.Dims()
)
shape_tensor = cat(
ctx,
target,
source_ir,
name + "_shape_concat",
tuple(shapes),
0,
cast_dtype=trt.int32,
)
start_tensor = cat(
ctx,
target,
source_ir,
name + "_start_concat",
starts,
0,
cast_dtype=trt.int32,
)
stride_tensor = cat(
ctx,
target,
source_ir,
name + "_stride_concat",
strides,
0,
cast_dtype=trt.int32,
)
layer.set_input(1, start_tensor)
layer.set_input(2, shape_tensor)
layer.set_input(3, stride_tensor)
layer.mode = trt.SampleMode.WRAP
set_layer_name(layer, target, name)
return layer.get_output(0)
Expand Down
45 changes: 45 additions & 0 deletions tests/py/dynamo/conversion/test_tile_aten.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@
import torch.nn as nn
from parameterized import parameterized
from torch.testing._internal.common_utils import run_tests
from torch_tensorrt import Input

from .harness import DispatchTestCase

Expand Down Expand Up @@ -71,5 +72,49 @@ def forward(self, x):
)


class TestTileConverterDynamicShape(DispatchTestCase):
@parameterized.expand(
[
((3,), (3,), (6,), (1,)),
((3,), (3,), (6,), (0,)),
((3,), (3,), (6,), (2,)),
((2,), (3,), (6,), (2, 2)),
((2,), (3,), (6,), (0, 2)),
# 2d cases
((3, 1), (3, 1), (6, 1), (0,)),
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is tiling with dims = (0,) same as tiling with (1,) ?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Tiling with dims = (0,) results in that particular dimension as 0 in the tiled tensor. In the above case though it prepends -1 resulting in dims being (-1, 0) post which the dynamic input tensor is tiled

((3, 1), (3, 1), (6, 1), (2,)),
((2, 3), (2, 3), (4, 3), (2, 2)),
((2, 3), (2, 3), (4, 3), (1, 0)),
((2, 3), (2, 3), (4, 3), (0, 2)),
((2, 3), (2, 3), (4, 3), (4, 2, 3)),
((2, 3), (2, 3), (4, 3), (0, 0, 3)),
((2, 3), (2, 3), (4, 3), (4, 2, 3, 1, 2)),
# 3d cases
((4, 2, 3), (4, 2, 3), (6, 2, 3), (2,)),
((4, 2, 3), (4, 2, 3), (6, 2, 3), (1, 2)),
((1, 2, 3), (1, 2, 3), (6, 2, 3), (2, 3)),
((1, 2, 3), (1, 2, 3), (6, 2, 3), (2, 3, 4)),
((1, 2, 3), (1, 2, 3), (6, 2, 3), (2, 3, 4, 5)),
]
)
def test_tile_input_dynamic(self, min_shape, opt_shape, max_shape, dims):
class Tile(nn.Module):
def forward(self, x):
return torch.ops.aten.tile.default(x, dims)

input_specs = [
Input(
min_shape=min_shape,
opt_shape=opt_shape,
max_shape=max_shape,
dtype=torch.float32,
),
]
self.run_test_with_dynamic_shape(
Tile(),
input_specs,
)


if __name__ == "__main__":
run_tests()