Skip to content

[TOPI][CUDA] Fix topk/sort gridDim overflow by remapping grid axes in sort_ir #19900

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

Open
cchung100m wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[TOPI][CUDA] Fix topk/sort gridDim overflow by remapping grid axes in sort_ir #19900

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
26cddfa
[TOPI][CUDA] Fix topk/sort gridDim.y overflow by moving the width-blo…
cchung100m Jun 28, 2026
b84daff
[TOPI][CUDA] Fix topk/sort gridDim overflow by remapping grid axes in…
cchung100m Jun 29, 2026
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
10 changes: 5 additions & 5 deletions python/tvm/topi/gpu/sort.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -499,17 +499,17 @@ def dual_mergepath(
nbx = tvm.tirx.generic.cast(ceil_div(width, max_threads * thread_work), "int32")
nbz = tvm.tirx.generic.cast(ceil_div(size, width), "int32")

tx, bx, by, _, _, _ = _get_threads(ntx, nbx, nthread_by * nbz)
tx, bx, by, _, _, _ = _get_threads(ntx, nbx, nthread_by)
bz = te.thread_axis("blockIdx.z")
with T.frame_scope(
[
T.attr(tx, "thread_extent", ntx),
T.attr(bx, "thread_extent", nbx),
T.attr(by, "thread_extent", nthread_by * nbz),
T.attr(by, "thread_extent", nthread_by),
T.attr(bz, "thread_extent", nbz),
]

@gemini-code-assist gemini-code-assist Bot Jun 28, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

high

In CUDA, the maximum limit for blockIdx.y and blockIdx.z is 65,535, while blockIdx.x can go up to $2^{31} - 1$ (2,147,483,647).

With the current changes in this PR:

  • bx (extent nbx) is bound to blockIdx.x
  • by (extent nthread_by) is bound to blockIdx.y
  • bz (extent nbz) is bound to blockIdx.z

Since nbz = ceil_div(size, width) and width starts at block_size (128), at the beginning of the merge sort, nbz can be up to ceil_div(size, 128). If size exceeds 128 * 65,535 = 8,388,480 (approx 8.38 million elements), nbz will exceed 65,535, causing a blockIdx.z overflow and kernel launch failure.

To support much larger sizes without overflow, we can swap the mapping of bx and bz:

  • Bind bz (extent nbz) to blockIdx.x (limit 2,147,483,647)
  • Bind bx (extent nbx) to blockIdx.z (limit 65,535)

Since nbx = ceil_div(width, max_threads * thread_work) <= ceil_div(size, 4096), nbx will only exceed 65,535 if size exceeds 4096 * 65,535 = 268,435,456 (approx 268 million elements). This simple swap increases the maximum supported sort size by 32x (from 8.38M to 268M elements) without adding any division/modulo overhead!

Suggested change
tx, bx, by, _, _, _ = _get_threads(ntx, nbx, nthread_by)
bz = te.thread_axis("blockIdx.z")
with T.frame_scope(
[
T.attr(tx, "thread_extent", ntx),
T.attr(bx, "thread_extent", nbx),
T.attr(by, "thread_extent", nthread_by * nbz),
T.attr(by, "thread_extent", nthread_by),
T.attr(bz, "thread_extent", nbz),
]
tx = te.thread_axis("threadIdx.x")
bx = te.thread_axis("blockIdx.z")
by = te.thread_axis("blockIdx.y")
bz = te.thread_axis("blockIdx.x")
with T.frame_scope(
[
T.attr(tx, "thread_extent", ntx),
T.attr(bx, "thread_extent", nbx),
T.attr(by, "thread_extent", nthread_by),
T.attr(bz, "thread_extent", nbz),
]
):

):
by_val = by % nthread_by
bz = by // nthread_by
base_idx = by_val * size
base_idx = by * size

# calculate the start, mid, and end points of this section
start_pos = width * bz
Expand Down
Loading