Skip to content

[ET-VK] Avoid the near-square local work group on Adreno #22328

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
msluszniak wants to merge 1 commit into
base: main
Choose a base branch
from
+11 −1
Open

[ET-VK] Avoid the near-square local work group on Adreno #22328

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
12 changes: 11 additions & 1 deletion backends/vulkan/runtime/graph/ops/impl/Common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,6 +132,15 @@ LocalWorkGroup pick_required_lwg(
return gwg.required_lwg_size();
}

// Adreno drivers miscompute the tiled GEMM family of shaders for near-square
// local work groups. With an 8x8x1 group (and with 4x16x1) entire work groups
// intermittently write garbage while every other block is bit-exact. These
// shaders have no shared memory and no barriers and their invocations are
// independent, so the result cannot legitimately depend on the group shape.
// A 2:1 x:y ratio, which yields 16x4x1 at the default 64 threads, is correct
// across every run measured and is the cheapest correct shape. See #22327.
const LwgShape kWideXLwg{2u, 1u, 0u};

LocalWorkGroup pick_xy_square_lwg(
ComputeGraph* graph,
const vkapi::ShaderInfo& shader,
Expand All @@ -142,7 +151,8 @@ LocalWorkGroup pick_xy_square_lwg(
(void)args;
(void)resize_args;
LocalWorkGroup lwg(
kSquareLwg, graph->context()->adapter_ptr()->recommended_lwg_nthreads());
graph->device_is_adreno() ? kWideXLwg : kSquareLwg,
graph->context()->adapter_ptr()->recommended_lwg_nthreads());
lwg.fit_to_global(gwg);
return lwg;
}
Expand Down
Loading