Skip to content

[WIP] [StreamK] #782

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

Draft
zhanglx13 wants to merge 2 commits into
base: main_perf
Choose a base branch
from
Draft

[WIP] [StreamK] #782

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c243ada
Set SM and kpack according to arch and early return if no streamK tiles
zhanglx13 Apr 27, 2025
6f09d02
Enable loop fusion
zhanglx13 Apr 29, 2025
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
5 changes: 3 additions & 2 deletions python/perf-kernels/streamk/03-matrix-multiplication-stream-k.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,7 +156,8 @@ def forward(ctx, a: torch.Tensor, b: torch.Tensor, c: torch.Tensor, bias: torch.

## test for tiles that is not multipe of 304 tiles
#m, n, k = 4096, 4096, 8192 # some problem size to test
m, n, k = 8192, 8192, 8192 # some problem size to test
m, n, k = 8192, 8192, 512 # some problem size to test
#m, n, k = 8704, 8704, 8192 # some problem size to test
#m, n, k = 512, 512, 512 # some problem size to test

## memory bound sizes
Expand Down Expand Up @@ -220,7 +221,7 @@ def forward(ctx, a: torch.Tensor, b: torch.Tensor, c: torch.Tensor, bias: torch.
P = torch.zeros((total_sm, BLK_M * BLK_N), device="cuda", dtype=torch.float32)
C = matmul.apply(A, B, C, bias, P, locks, total_sm, BLK_M, BLK_N, BLK_K, gsize_m, two_tiles, num_stages, num_warps,
waves_per_eu, mfmaInstrSize, kpack)
#exit(0)
exit(0)
matmul.set_debug(False)
expected = A @ B

Expand Down
6 changes: 4 additions & 2 deletions python/perf-kernels/streamk/streamk_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,7 @@ def streamk_gemm(

acc_dtype = tl.float32 if C.type.element_ty != tl.int8 else tl.int32

for tile_id in range(pid, total_full_tiles, NUM_SMS):
for tile_id in tl.range(pid, total_full_tiles, NUM_SMS, flatten=True):
Copy link
Member

@xiaohuguo2023 xiaohuguo2023 Apr 29, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

have you observed any perf improvement by enable this loop fusion ?

Copy link
Author

@zhanglx13 zhanglx13 Apr 29, 2025

Choose a reason for hiding this comment

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

no, this is just an experiment and it doesn't compile. I filed a ticket for it https://github.com/ROCm/triton-internal/issues/784

num_pid_in_group = GROUP_SIZE_M * num_pid_n
group_id = tile_id // num_pid_in_group
first_pid_m = group_id * GROUP_SIZE_M
Expand All @@ -74,8 +74,10 @@ def streamk_gemm(
if not EVEN_K:
loop_k -= 1

tl.assume(loop_k > 1)

Copy link
Member

@xiaohuguo2023 xiaohuguo2023 Apr 29, 2025

Choose a reason for hiding this comment

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

nice catch!

acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=acc_dtype)
for k in range(0, loop_k):
for k in tl.range(0, loop_k):
a = tl.load(tl.multiple_of(A_BASE, (1, 16)))
b = tl.load(tl.multiple_of(B_BASE, (16, 1)))
acc += tl.dot(a, b)
Expand Down