Skip to content

[Triton/Gluon] enable TDM fusion on MXFP4 #6097

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
Boss2002n merged 1 commit into from
Oct 2, 2026
Merged

[Triton/Gluon] enable TDM fusion on MXFP4 #6097

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
60 changes: 47 additions & 13 deletions aiter/ops/triton/_gluon_kernels/gfx1250/gemm/basic/gemm_mxfp4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,8 +47,10 @@ def get_gemm_afp4wfp4_preshuffle_layouts(num_warps, BLOCK_M, BLOCK_N, BLOCK_K):
shared_B = gl.SwizzledSharedLayout(vec=1, per_phase=1, max_phase=1, order=[1, 0])
shared_S = gl.SwizzledSharedLayout(vec=1, per_phase=1, max_phase=1, order=[1, 0])

# Output staging layout for the TDM store (acc -> LDS -> HBM)
shared_C = gl.SwizzledSharedLayout(vec=1, per_phase=1, max_phase=1, order=[1, 0])
# Output staging layout for the TDM store (acc -> LDS -> HBM).
shared_C = gl.PaddedSharedLayout.with_identity_for(
[[BLOCK_N, 8]], [BLOCK_M, BLOCK_N], [1, 0]
)
Comment thread
Boss2002n marked this conversation as resolved.

# Register layouts for WMMA operands
dot_a = gl.DotOperandLayout(operand_index=0, parent=wmma_layout, k_width=16)
Expand Down Expand Up @@ -130,6 +132,37 @@ def depreshuffle_b_raw_to_kn(
)


@gluon.jit
def _issue_tdm_stage(
a_desc,
b_desc,
as_desc,
bs_desc,
a_slot,
b_slot,
as_slot,
bs_slot,
off_a,
off_b,
off_as,
off_bs,
):
# Movbe all descriptors before the four copies then the compiler fuses all TDM copies
Comment thread
Boss2002n marked this conversation as resolved.
a_desc = gl.amd.gfx1250.tdm.update_tensor_descriptor(a_desc, add_offsets=[0, off_a])
b_desc = gl.amd.gfx1250.tdm.update_tensor_descriptor(b_desc, add_offsets=[0, off_b])
as_desc = gl.amd.gfx1250.tdm.update_tensor_descriptor(
as_desc, add_offsets=[0, off_as]
)
bs_desc = gl.amd.gfx1250.tdm.update_tensor_descriptor(
bs_desc, add_offsets=[0, off_bs]
)

gl.amd.gfx1250.tdm.async_load(a_desc, dest=a_slot)
gl.amd.gfx1250.tdm.async_load(b_desc, dest=b_slot)
gl.amd.gfx1250.tdm.async_load(as_desc, dest=as_slot)
gl.amd.gfx1250.tdm.async_load(bs_desc, dest=bs_slot)


_gemm_mxfp4_preshuffle_gfx1250_repr = make_kernel_repr(
"_gemm_mxfp4_preshuffle_gfx1250_kernel",
[
Expand Down Expand Up @@ -179,6 +212,13 @@ def gemm_mxfp4_preshuffle_gfx1250(
a_scale_layout: gl.constexpr,
b_scale_layout: gl.constexpr,
):
# async_wait counts TDM ops in flight. The compiler fuses each stage's four
# copies into one op (two with 2 warps, which fuse in pairs).
if num_warps == 2:
TDM_OPS_PER_STAGE: gl.constexpr = 2
else:
TDM_OPS_PER_STAGE: gl.constexpr = 1

# Compile-time constants
FP4_ELEMS_PER_BYTE: gl.constexpr = 2
SCALE_GROUP_ELEMS: gl.constexpr = 32
Expand Down Expand Up @@ -315,14 +355,11 @@ def gemm_mxfp4_preshuffle_gfx1250(
off_b = load_idx * BLOCK_K_BYTES * 16
off_as = load_idx * K_GROUPS * A_PRESHUFFLE_FACTOR
off_bs = load_idx * K_GROUPS * PRESHUFFLE_FACTOR
gl.amd.gfx1250.tdm.async_load(a_desc, [0, off_a], a_slot)
gl.amd.gfx1250.tdm.async_load(b_desc, [0, off_b], b_slot)
gl.amd.gfx1250.tdm.async_load(as_desc, [0, off_as], as_slot)
gl.amd.gfx1250.tdm.async_load(bs_desc, [0, off_bs], bs_slot)
_issue_tdm_stage(a_desc, b_desc, as_desc, bs_desc, a_slot, b_slot, as_slot, bs_slot, off_a, off_b, off_as, off_bs) # fmt: skip
load_idx += 1

# --- 2. Pre-load tile 0 from LDS into registers ---
gl.amd.gfx1250.tdm.async_wait((NUM_BUFFERS - 1) * 4)
gl.amd.gfx1250.tdm.async_wait((NUM_BUFFERS - 1) * TDM_OPS_PER_STAGE)

slot_c = compute_idx % NUM_BUFFERS
cur_A = smem_A.index(slot_c).load(layout=dot_a_layout)
Expand Down Expand Up @@ -355,12 +392,9 @@ def gemm_mxfp4_preshuffle_gfx1250(
off_b = load_idx * BLOCK_K_BYTES * 16
off_as = load_idx * K_GROUPS * A_PRESHUFFLE_FACTOR
off_bs = load_idx * K_GROUPS * PRESHUFFLE_FACTOR
gl.amd.gfx1250.tdm.async_load(a_desc, [0, off_a], a_slot)
gl.amd.gfx1250.tdm.async_load(b_desc, [0, off_b], b_slot)
gl.amd.gfx1250.tdm.async_load(as_desc, [0, off_as], as_slot)
gl.amd.gfx1250.tdm.async_load(bs_desc, [0, off_bs], bs_slot)
_issue_tdm_stage(a_desc, b_desc, as_desc, bs_desc, a_slot, b_slot, as_slot, bs_slot, off_a, off_b, off_as, off_bs) # fmt: skip

gl.amd.gfx1250.tdm.async_wait((NUM_BUFFERS - 1) * 4)
gl.amd.gfx1250.tdm.async_wait((NUM_BUFFERS - 1) * TDM_OPS_PER_STAGE)
load_idx += 1

# Pre-load next tile from LDS into registers
Expand Down Expand Up @@ -393,7 +427,7 @@ def gemm_mxfp4_preshuffle_gfx1250(
cur_A, cur_AS, "e2m1", cur_B, cur_BS, "e2m1", acc
)

gl.amd.gfx1250.tdm.async_wait((NUM_BUFFERS - 2 - i) * 4)
gl.amd.gfx1250.tdm.async_wait((NUM_BUFFERS - 2 - i) * TDM_OPS_PER_STAGE)

next_slot = (compute_idx + 1) % NUM_BUFFERS
cur_A = smem_A.index(next_slot).load(layout=dot_a_layout)
Expand Down
Loading