Skip to content

[Mosaic GPU][NFC] Derive smem_buffer_ptr in OnDeviceProfiler.record. #32442

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

Merged
merged 1 commit into from
Oct 10, 2025
+8 −9
Merged

[Mosaic GPU][NFC] Derive smem_buffer_ptr in OnDeviceProfiler.record. #32442

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
17 changes: 8 additions & 9 deletions jax/experimental/mosaic/gpu/profiler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -315,31 +315,30 @@ def __init__(self, spec: ProfilerSpec, smem_buffer: ir.Value, gmem_buffer: ir.Va
self.entries_per_wg,
),
)
self.smem_buffer_ptr = memref_ptr(self.smem_buffer, memory_space=3)
self.gmem_buffer = gmem_buffer
self.is_profiling_thread = arith.cmpi(
arith.CmpIPredicate.eq,
arith.remui(thread_idx(), c(WARPGROUP_SIZE, i32)),
c(0, i32),
)
# Hopefully mem2reg will remove the allocation.
self.offset = memref.alloca(ir.MemRefType.get((), i32), [], [])
memref.store(c(0, i32), self.offset, [])
self.offset = memref.alloca(ir.MemRefType.get((), index), [], [])
memref.store(c(0, index), self.offset, [])

@contextlib.contextmanager
def record(self, name: str):
i32 = ir.IntegerType.get_signless(32)
index = ir.IndexType.get()
name_id = self.spec.intern_name(name)
def store(modifier):
# smem_buffer[offset] = modifier | name_id
# smem_buffer[offset + 1] = %clock
# offset += 2
offset = memref.load(self.offset, [])
base_ref = memref_slice(self.smem_buffer, offset)
base_ptr = memref_ptr(base_ref, memory_space=3)
i64 = ir.IntegerType.get_signless(64)
base_addr = arith.addi(
llvm.ptrtoint(i64, self.smem_buffer_ptr),
arith.extui(i64, arith.muli(offset, c(4, i32))),
)
base_addr = llvm.ptrtoint(i64, base_ptr)
llvm.inline_asm(
ir.Type.parse("!llvm.void"),
[self.is_profiling_thread, base_addr, c(modifier | name_id, i32)],
Expand All @@ -349,7 +348,7 @@ def store(modifier):
"b,l,r",
has_side_effects=True,
)
new_offset = arith.addi(offset, c(2, i32))
new_offset = arith.addi(offset, c(2, index))
memref.store(new_offset, self.offset, [])
store(ProfilerSpec.ENTER)
yield
Expand Down Expand Up @@ -379,7 +378,7 @@ def finalize(self, grid: tuple[int, ...], block: tuple[int, ...]):
with when(self.is_profiling_thread):
memref.store(self.start, wg_gmem_buffer, [c(0, index)])
memref.store(smid(), wg_gmem_buffer, [c(1, index)])
num_traces = memref.load(self.offset, [])
num_traces = arith.index_cast(i32, memref.load(self.offset, []))
memref.store(num_traces, wg_gmem_buffer, [c(2, index)])
traces = vector.load(
ir.VectorType.get((self.entries_per_wg - 3,), i32),
Expand Down
Loading