Skip to content

[Relay][Strategy] Allow cuda cross compilation without physical device. #7063

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
comaniac merged 5 commits into from
Dec 10, 2020
Merged

[Relay][Strategy] Allow cuda cross compilation without physical device. #7063

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
fe49659
Allow cross compilation of cuda targets without physical device.
Dec 8, 2020
d7f45dc
Formatting.
Dec 8, 2020
6af0714
Add warning when architecture cant be found.
jwfromm Dec 9, 2020
cdd7daf
Use target instead of autotvm arch specification.
Dec 9, 2020
deee3c3
Change warning message.
Dec 9, 2020
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
25 changes: 22 additions & 3 deletions python/tvm/contrib/nvcc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -269,15 +269,34 @@ def have_int8(compute_version):
return False


def have_tensorcore(compute_version):
def have_tensorcore(compute_version=None, target=None):
"""Either TensorCore support is provided in the compute capability or not

Parameters
----------
compute_version : str
compute capability of a GPU (e.g. "7.0")
compute_version : str, optional
compute capability of a GPU (e.g. "7.0").

target : tvm.target.Target, optional
The compilation target, will be used to determine arch if compute_version
isn't specified.
"""
if compute_version is None:
if tvm.gpu(0).exist:
compute_version = tvm.gpu(0).compute_version
else:
if target is None or "arch" not in target.attrs:
warnings.warn(
"Tensorcore will be disabled due to no CUDA architecture specified."
"Try specifying it by adding '-arch=sm_xx' to your target."
)
return False
compute_version = target.attrs["arch"]
# Compute version will be in the form "sm_{major}{minor}"
major, minor = compute_version.split("_")[1]
compute_version = major + "." + minor
major, _ = parse_compute_version(compute_version)

if major == 7:
return True

Expand Down
10 changes: 5 additions & 5 deletions python/tvm/relay/op/strategy/cuda.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,7 +197,7 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
if judge_winograd_autotvm:
if (
target.kind.name == "cuda"
and nvcc.have_tensorcore(tvm.gpu(0).compute_version)
and nvcc.have_tensorcore(target=target)
and judge_winograd_tensorcore
):
strategy.add_implementation(
Expand All @@ -215,7 +215,7 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
)
if (
target.kind.name == "cuda"
and nvcc.have_tensorcore(tvm.gpu(0).compute_version)
and nvcc.have_tensorcore(target=target)
and (
(N % 16 == 0 and CI % 16 == 0 and CO % 16 == 0)
or (N % 8 == 0 and CI % 16 == 0 and CO % 32 == 0)
Expand Down Expand Up @@ -436,7 +436,7 @@ def conv2d_winograd_without_weight_transfrom_strategy_cuda(attrs, inputs, out_ty
)
if (
target.kind.name == "cuda"
and nvcc.have_tensorcore(tvm.gpu(0).compute_version)
and nvcc.have_tensorcore(target=target)
and judge_winograd_tensorcore
):
strategy.add_implementation(
Expand Down Expand Up @@ -563,7 +563,7 @@ def conv3d_strategy_cuda(attrs, inputs, out_type, target):
N, _, _, _, _ = get_const_tuple(data.shape)
_, _, _, CI, CO = get_const_tuple(kernel.shape)
if target.kind.name == "cuda":
if nvcc.have_tensorcore(tvm.gpu(0).compute_version):
if nvcc.have_tensorcore(target=target):
if (
(N % 16 == 0 and CI % 16 == 0 and CO % 16 == 0)
or (N % 8 == 0 and CI % 16 == 0 and CO % 32 == 0)
Expand Down Expand Up @@ -679,7 +679,7 @@ def dense_strategy_cuda(attrs, inputs, out_type, target):
plevel=5,
)
if target.kind.name == "cuda":
if nvcc.have_tensorcore(tvm.gpu(0).compute_version):
if nvcc.have_tensorcore(target=target):
if (
(i % 16 == 0 and b % 16 == 0 and o % 16 == 0)
or (i % 16 == 0 and b % 8 == 0 and o % 32 == 0)
Expand Down