Skip to content

[ROCM] Fixes compiling on ROCM 5 and accuracy on dense op #13847

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
junrushao merged 1 commit into from
Jan 26, 2023
Merged

[ROCM] Fixes compiling on ROCM 5 and accuracy on dense op #13847

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
4 changes: 2 additions & 2 deletions python/tvm/contrib/rocm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -141,9 +141,7 @@ def callback_rocm_bitcode_path(rocdl_dir=None):
bitcode_names = [
"oclc_daz_opt_on",
"ocml",
"hc",
"irif", # this does not exist in rocm 3.9, drop eventually
"ockl",
"oclc_correctly_rounded_sqrt_off",
"oclc_correctly_rounded_sqrt_on",
"oclc_daz_opt_off",
Expand All @@ -152,9 +150,11 @@ def callback_rocm_bitcode_path(rocdl_dir=None):
"oclc_isa_version_803", # todo (t-vi): an alternative might be to scan for the
"oclc_isa_version_900", # isa version files (if the linker throws out
"oclc_isa_version_906", # the unneeded ones or we filter for the arch we need)
"oclc_isa_version_1030",
"oclc_unsafe_math_off",
"oclc_unsafe_math_on",
"oclc_wavefrontsize64_on",
"oclc_abi_version_500",
]

bitcode_files = []
Expand Down
15 changes: 9 additions & 6 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 @@ -918,13 +918,16 @@ def dense_strategy_cuda(attrs, inputs, out_type, target):
name="dense_int8.cuda",
)
else:
strategy.add_implementation(
wrap_compute_dense(topi.gpu.dense_small_batch),
wrap_topi_schedule(topi.gpu.schedule_dense_small_batch),
name="dense_small_batch.gpu",
)
# Some AMDGPU cards have accuracy issues with this schedule
# See https://github.com/apache/tvm/issues/13666
if target.kind.name != "rocm":
strategy.add_implementation(
wrap_compute_dense(topi.gpu.dense_small_batch),
wrap_topi_schedule(topi.gpu.schedule_dense_small_batch),
name="dense_small_batch.gpu",
)

with SpecializedCondition(b >= 32):
with SpecializedCondition(target.kind.name == "rocm" or b >= 32):
strategy.add_implementation(
wrap_compute_dense(topi.gpu.dense_large_batch),
wrap_topi_schedule(topi.gpu.schedule_dense_large_batch),
Expand Down