Skip to content

Add missing preamble functions to reset locks in examples 10 and 11 #213

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
wants to merge 11 commits into
base: main
Choose a base branch
from
Open

Add missing preamble functions to reset locks in examples 10 and 11 #213

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
020578f
Initial plan
Copilot Oct 9, 2025
16ce1d8
Add preamble function to reset locks in examples 10 and 11
Copilot Oct 9, 2025
f585cae
Merge branch 'main' into copilot/fix-missing-preamble-examples
mawad-amd Oct 9, 2025
8d15de7
Use proper synchronization instead of loads and stores (#215)
mawad-amd Oct 9, 2025
a5ec30c
Fix atomic operations: use atomic_cas for non-destructive polling
Copilot Oct 10, 2025
f32e69a
Fix atomic_cas polling: use (1, 0) to atomically consume the flag
Copilot Oct 10, 2025
7479b4b
Use volatile load spin loop followed by atomic_cas with dependency
Copilot Oct 10, 2025
c66903c
Add port cleanup to CI tests to prevent lingering connections
Copilot Oct 10, 2025
7d7179b
Revert "Add port cleanup to CI tests to prevent lingering connections"
Copilot Oct 10, 2025
acc2760
Merge branch 'main' into copilot/fix-missing-preamble-examples
mawad-amd Oct 10, 2025
b501b8c
Merge branch 'main' into copilot/fix-missing-preamble-examples
mawad-amd Oct 10, 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
9 changes: 7 additions & 2 deletions examples/10_gemm_all_scatter_wg_specialization/benchmark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,7 +136,7 @@ def _worker(local_rank: int, world_size: int, init_url: str, args: dict):
total_blocks_N = triton.cdiv(args["n"], args["BLK_N"])
total_tiles = total_blocks_M * total_blocks_N

locks = shmem.zeros((total_tiles,), device="cuda", dtype=torch.int8)
locks = shmem.zeros((total_tiles,), device="cuda", dtype=torch.int32)

bias = None

Expand All @@ -157,6 +157,11 @@ def _worker(local_rank: int, world_size: int, init_url: str, args: dict):
# Allocate Timestamps
timestamps = Timestamps(num_tiles=total_tiles)

def preamble():
shmem.barrier()
locks.zero_()
shmem.barrier()

def run_experiment():
nonlocal local_C
nonlocal global_C
Expand Down Expand Up @@ -244,7 +249,7 @@ def run_experiment():
matmul.set_debug(False)
shmem.info("Benchmarking...")
perf = lambda ms: 2 * args["M"] * args["N"] * args["K"] * 1e-12 / (ms * 1e-3)
triton_ms = iris.do_bench(run_experiment, shmem.barrier)
triton_ms = iris.do_bench(run_experiment, shmem.barrier, preamble)
triton_tflops = perf(triton_ms)
algo_string = "all_scatter"
shmem.info(
Expand Down
5 changes: 2 additions & 3 deletions examples/10_gemm_all_scatter_wg_specialization/gemm_all_scatter_wg_specialization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -142,8 +142,7 @@ def persistent_gemm_all_scatter_wg_specialization(
tl.atomic_max(mm_end_timestamp_ptr + tile_id, timestamp)

tl.store(c_global + global_offset, c, mask=sub_mask, cache_modifier=".wt")
tl.debug_barrier()
tl.store(locks + tile_id, 1, cache_modifier=".wt")
tl.atomic_xchg(locks + tile_id, 1, sem="release", scope="gpu")

else: # pid >= GEMM_SMS
COMM_SMS = NUM_SMS - GEMM_SMS
Expand All @@ -165,7 +164,7 @@ def persistent_gemm_all_scatter_wg_specialization(
global_offset = rm[:, None] * stride_cm_global + (rn[None, :] + cur_rank * N) * stride_cn_global
# End: masks/offset calculations.

while tl.load(locks + tile_id, cache_modifier=".cv", volatile=True) != 1:
while tl.atomic_xchg(locks + tile_id, 0, sem="acquire", scope="gpu") != 1:
pass

for remote_rank in range(world_size):
Expand Down
9 changes: 7 additions & 2 deletions examples/11_gemm_all_scatter_producer_consumer/benchmark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,7 +136,7 @@ def _worker(local_rank: int, world_size: int, init_url: str, args: dict):
total_blocks_N = triton.cdiv(args["n"], args["BLK_N"])
total_tiles = total_blocks_M * total_blocks_N

locks = shmem.zeros((total_tiles,), device="cuda", dtype=torch.int8)
locks = shmem.zeros((total_tiles,), device="cuda", dtype=torch.int32)

bias = None

Expand Down Expand Up @@ -166,6 +166,11 @@ def _worker(local_rank: int, world_size: int, init_url: str, args: dict):
# Allocate Timestamps
timestamps = Timestamps(num_tiles=total_tiles)

def preamble():
shmem.barrier()
locks.zero_()
shmem.barrier()

def run_experiment():
nonlocal C
nonlocal kernel_timing
Expand Down Expand Up @@ -275,7 +280,7 @@ def run_experiment():
matmul.set_debug(False)
shmem.info("Benchmarking...")
perf = lambda ms: 2 * args["M"] * args["N"] * args["K"] * 1e-12 / (ms * 1e-3)
triton_ms = iris.do_bench(run_experiment, shmem.barrier)
triton_ms = iris.do_bench(run_experiment, shmem.barrier, preamble)
triton_tflops = perf(triton_ms)
algo_string = "all_scatter"
shmem.info(
Expand Down
5 changes: 2 additions & 3 deletions examples/11_gemm_all_scatter_producer_consumer/gemm_all_scatter_producer_consumer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,8 +133,7 @@ def persistent_gemm(
tl.atomic_max(mm_end_timestamp_ptr + tile_id, timestamp)

tl.store(C + global_offset, c, mask=sub_mask, cache_modifier=".wt")
tl.debug_barrier()
tl.store(locks + tile_id, 1, cache_modifier=".wt")
tl.atomic_xchg(locks + tile_id, 1, sem="release", scope="gpu")


@triton.jit()
Expand Down Expand Up @@ -185,7 +184,7 @@ def persistent_all_scatter(
global_offset = rm[:, None] * stride_cm_global + (rn[None, :] + cur_rank * N) * stride_cn_global
# End: masks/offset calculations.

while tl.load(locks + tile_id, cache_modifier=".cv", volatile=True) != 1:
while tl.atomic_xchg(locks + tile_id, 0, sem="acquire", scope="gpu") != 1:
pass

for remote_rank in range(world_size):
Expand Down
Loading