Skip to content

[DLight] Perf improvement for low_batch_gemv on Metal #17026

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
tqchen merged 1 commit into from
May 24, 2024
Merged

[DLight] Perf improvement for low_batch_gemv on Metal #17026

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
13 changes: 6 additions & 7 deletions python/tvm/dlight/gpu/low_batch_gemv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -500,7 +500,7 @@ def apply(
sch.set_scope(block, 0, "shared")
_, _, _, *s = sch.get_loops(epilogue) # pylint: disable=invalid-name
_, tx = sch.split(sch.fuse(*s), factors=[None, TX])
sch.bind(tx, "threadIdx.x")
sch.bind(tx, TAG_S)
else:
sch.reverse_compute_at(epilogue, bx, preserve_unit_loops=True)
ts_tile_s = sch.fuse(*sch.get_loops(epilogue)[3:])
Expand Down Expand Up @@ -538,17 +538,16 @@ def apply(
else:
TS, TR = 16, 32
elif target.kind.name == "metal":
# Note that the following tile size is tuned on M2 Ultra for 7B
TAG_S, TAG_R = "threadIdx.x", "threadIdx.y"
VEC_C = 1
VEC_C = 4
LOAD_V_SHARED = False
LOAD_V_VEC = -1
UNROLL = 256
UNROLL = 8
if isinstance(len_S, int):
if len_S > len_R:
TS, TR = 2, 32
TS, TR = 8, 32
else:
TS, TR = 2, 64
TAG_S, TAG_R = "threadIdx.x", "threadIdx.y"
TS, TR = 8, 32
elif target.kind.name == "rocm":
VEC_C = 4
LOAD_V_SHARED = True
Expand Down
Loading