Skip to content

Fix TMEM address read/write race in example 77 #2835

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
milesvant wants to merge 1 commit into
base: main
Choose a base branch
from
+20 −20
Draft

Fix TMEM address read/write race in example 77 #2835

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
40 changes: 20 additions & 20 deletions examples/77_blackwell_fmha/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -468,7 +468,7 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
else if (role == WarpRole::Correction) {
cutlass::arch::warpgroup_reg_dealloc<NumRegsCorrection>();

bool has_valid = false;
bool tmem_allocated = false;

CUTLASS_PRAGMA_NO_UNROLL
for (; tile_scheduler.is_valid(); ++tile_scheduler) {
Expand All @@ -481,8 +481,6 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
continue;
}

has_valid = true;

if (get<1>(logical_problem_shape) == 0) {
mainloop.correction_empty(
blk_coord,
Expand All @@ -495,6 +493,8 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
continue;
}

tmem_allocated = true;

mainloop.correction(
blk_coord,
params.mainloop, logical_problem_shape,
Expand All @@ -512,7 +512,7 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
if constexpr (NumWarpsEpilogue == 0) {
static_assert(NumWarpsCorrection == 1);

if (has_valid) {
if (tmem_allocated) {
uint32_t free_stage_ptr = shared_storage.tmem_base_ptr;
tmem_allocator.free(free_stage_ptr, TmemAllocator::Sm100TmemCapacityColumns);
}
Expand All @@ -522,7 +522,7 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
else if (role == WarpRole::MMA) {
warpgroup_reg_set<NumRegsOther>();

bool allocated = false;
bool tmem_allocated = false;

CUTLASS_PRAGMA_NO_UNROLL
for (; tile_scheduler.is_valid(); ++tile_scheduler) {
Expand All @@ -531,18 +531,17 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
auto logical_problem_shape = apply_batch(params,
params.problem_shape, get<2,1>(blk_coord));

if (get<0>(blk_coord) * get<0>(TileShape{}) >= get<0>(logical_problem_shape)) {
if (
(get<0>(blk_coord) * get<0>(TileShape{}) >= get<0>(logical_problem_shape))
|| (get<1>(logical_problem_shape) == 0)
) {
continue;
}

if (!allocated) {
Copy link
Contributor Author

@milesvant milesvant Dec 3, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

Reorders this TMEM allocation with the next continue statement

if (!tmem_allocated) {
tmem_allocator.allocate(TmemAllocator::Sm100TmemCapacityColumns, &shared_storage.tmem_base_ptr);
__syncwarp();
allocated = true;
}

if (get<1>(logical_problem_shape) == 0) {
continue;
tmem_allocated = true;
}

mainloop.mma(
Expand Down Expand Up @@ -573,11 +572,10 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
auto logical_problem_shape = apply_batch(params,
params.problem_shape, get<2,1>(blk_coord));

if (get<0>(blk_coord) * get<0>(TileShape{}) >= get<0>(logical_problem_shape)) {
continue;
}

if (get<1>(logical_problem_shape) == 0) {
if (
(get<0>(blk_coord) * get<0>(TileShape{}) >= get<0>(logical_problem_shape))
|| (get<1>(logical_problem_shape) == 0)
) {
continue;
}

Expand All @@ -594,7 +592,7 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
else if (role == WarpRole::Epilogue) {
warpgroup_reg_set<NumRegsOther>();

bool has_valid = false;
bool tmem_allocated = false;

CUTLASS_PRAGMA_NO_UNROLL
for (; tile_scheduler.is_valid(); ++tile_scheduler) {
Expand All @@ -607,7 +605,9 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {
continue;
}

has_valid = true;
if (get<1>(logical_problem_shape) != 0) {
tmem_allocated = true;
}

epilogue.store(
blk_coord, logical_problem_shape,
Expand All @@ -620,7 +620,7 @@ struct Sm100FmhaFwdKernelTmaWarpspecialized {

static_assert(NumWarpsEpilogue <= 1);
if constexpr (NumWarpsEpilogue == 1) {
if(has_valid) {
if(tmem_allocated) {
uint32_t free_stage_ptr = shared_storage.tmem_base_ptr;
tmem_allocator.free(free_stage_ptr, TmemAllocator::Sm100TmemCapacityColumns);
}
Expand Down