Skip to content

fix(sparse_mla): avoid immutable load-coordinate rebindings #1

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
ErvinXie wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(sparse_mla): avoid immutable load-coordinate rebindings #1

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
90 changes: 55 additions & 35 deletions mate/sparse_mla/tilelang/sparse_mla_v32_decode_fwd_scheduled.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -334,16 +334,16 @@ def dsa_decode_split(
kv_reg_l_bf16_load = T.alloc_local([32], T.bfloat16)
kv_reg_l_fp8 = T.view(kv_reg_l_bf16_load, [64], kv_latent_dtype)
quant_local_l = T.alloc_local([2, 2], T.float32)
ldg_tx = (tid) % 8
ldg_ty = (tid) // 8
c0_ldg_tx = (tid) % 8
c0_ldg_ty = (tid) // 8
T.fill(sumexp, 0)
T.fill(m_i, -(2**30))
T.fill(acc_o_l_0, 0)
T.fill(acc_o_l_1, 0)
for i_i in range(start_block_idx, end_block_idx):
T.barrier_wait(bar_kv0_ready, (phase_count[0] & 1))
T.copy(quant_shared[ldg_ty, 0:2], quant_local_l[0, :])
T.copy(quant_shared[ldg_ty + 32, 0:2], quant_local_l[1, :])
T.copy(quant_shared[c0_ldg_ty, 0:2], quant_local_l[0, :])
T.copy(quant_shared[c0_ldg_ty + 32, 0:2], quant_local_l[1, :])
T.annotate_layout(
{
kv_shared_l[
Expand All @@ -361,8 +361,8 @@ def dsa_decode_split(
for v in T.vectorized(4):
kv_reg_l_bf16_load[r * 16 + u * 4 + v] = (
kv_shared_l[
(ldg_ty + r * 32),
64 * u + ldg_tx * 8 + v,
(c0_ldg_ty + r * 32),
64 * u + c0_ldg_tx * 8 + v,
]
)
T.lma_wait()
Expand All @@ -386,16 +386,17 @@ def dsa_decode_split(
for u in T.unroll(2):
for v in T.vectorized(8):
kv_shared_l[
ldg_ty + r * 32, 64 * u + ldg_tx * 8 + v
c0_ldg_ty + r * 32,
64 * u + c0_ldg_tx * 8 + v,
] = kv_reg_l[r * 32 + u * 8 + v]

for r in T.unroll(2):
for u in T.unroll(2):
for v in T.vectorized(4):
kv_reg_l_bf16_load[r * 16 + (u + 2) * 4 + v] = (
kv_shared_l[
(ldg_ty + r * 32),
64 * (u + 2) + ldg_tx * 8 + v,
(c0_ldg_ty + r * 32),
64 * (u + 2) + c0_ldg_tx * 8 + v,
]
)
T.lma_wait()
Expand All @@ -418,7 +419,8 @@ def dsa_decode_split(
for u in T.unroll(2):
for v in T.vectorized(8):
kv_shared_l[
ldg_ty + r * 32, 64 * (u + 2) + ldg_tx * 8 + v
c0_ldg_ty + r * 32,
64 * (u + 2) + c0_ldg_tx * 8 + v,
] = kv_reg_l[r * 32 + (u + 2) * 8 + v]

T.lma_wait()
Expand Down Expand Up @@ -506,7 +508,9 @@ def dsa_decode_split(

T.lma_wait()
T.barrier_arrive(bar_p_ready)
stage_value_shared(v_shared_0, kv_reg_l, ldg_ty, ldg_tx, 0)
stage_value_shared(
v_shared_0, kv_reg_l, c0_ldg_ty, c0_ldg_tx, 0
)
T.lma_wait()
T.barrier_arrive(bar_vl0_ready)
T.barrier_wait(bar_vl0_ready, (phase_count[0] & 1))
Expand All @@ -519,7 +523,9 @@ def dsa_decode_split(
wg_wait=-1,
)
T.warpgroup_commit_batch()
stage_value_shared(v_shared_1, kv_reg_l, ldg_ty, ldg_tx, 2)
stage_value_shared(
v_shared_1, kv_reg_l, c0_ldg_ty, c0_ldg_tx, 2
)

T.warpgroup_wait(0)
T.barrier_arrive(bar_vl0_free)
Expand Down Expand Up @@ -576,12 +582,12 @@ def dsa_decode_split(
quant_local_r = T.alloc_local([2, 2], T.float32)
T.fill(acc_o_r_0, 0)
T.fill(acc_o_r_1, 0)
ldg_tx = (tid - 256) % 8
ldg_ty = (tid - 256) // 8
c1_ldg_tx = (tid - 256) % 8
c1_ldg_ty = (tid - 256) // 8
for i_i in range(start_block_idx, end_block_idx):
T.barrier_wait(bar_kv1_ready, (phase_count[0] & 1))
T.copy(quant_shared[ldg_ty, 2:4], quant_local_r[0, :])
T.copy(quant_shared[ldg_ty + 32, 2:4], quant_local_r[1, :])
T.copy(quant_shared[c1_ldg_ty, 2:4], quant_local_r[0, :])
T.copy(quant_shared[c1_ldg_ty + 32, 2:4], quant_local_r[1, :])

T.annotate_layout(
{
Expand All @@ -599,8 +605,8 @@ def dsa_decode_split(
for u in T.unroll(2):
for v in T.vectorized(4):
kv_reg_r_bf16[r * 16 + u * 4 + v] = kv_shared_r[
ldg_ty + r * 32,
64 * u + ldg_tx * 8 + v,
c1_ldg_ty + r * 32,
64 * u + c1_ldg_tx * 8 + v,
]
T.lma_wait()

Expand All @@ -623,16 +629,17 @@ def dsa_decode_split(
for u in T.unroll(2):
for v in T.vectorized(8):
kv_shared_r[
ldg_ty + r * 32, 64 * u + ldg_tx * 8 + v
c1_ldg_ty + r * 32,
64 * u + c1_ldg_tx * 8 + v,
] = kv_reg_r[r * 32 + u * 8 + v]

for r in T.unroll(2):
for u in T.unroll(2):
for v in T.vectorized(4):
kv_reg_r_bf16[r * 16 + (u + 2) * 4 + v] = (
kv_shared_r[
(ldg_ty + r * 32),
64 * (u + 2) + ldg_tx * 8 + v,
(c1_ldg_ty + r * 32),
64 * (u + 2) + c1_ldg_tx * 8 + v,
]
)
T.lma_wait()
Expand All @@ -656,15 +663,18 @@ def dsa_decode_split(
for u in T.unroll(2):
for v in T.vectorized(8):
kv_shared_r[
ldg_ty + r * 32, 64 * (u + 2) + ldg_tx * 8 + v
c1_ldg_ty + r * 32,
64 * (u + 2) + c1_ldg_tx * 8 + v,
] = kv_reg_r[r * 32 + (u + 2) * 8 + v]

T.lma_wait()
T.barrier_arrive(bar_kv1_quant_ready)

T.barrier_wait(bar_vl0_free, (phase_count[0] & 1))

stage_value_shared(v_shared_0, kv_reg_r, ldg_ty, ldg_tx, 0)
stage_value_shared(
v_shared_0, kv_reg_r, c1_ldg_ty, c1_ldg_tx, 0
)

T.lma_wait()
T.barrier_arrive(bar_vr0_ready)
Expand All @@ -688,7 +698,9 @@ def dsa_decode_split(

T.barrier_wait(bar_vl1_free, (phase_count[0] & 1))

stage_value_shared(v_shared_1, kv_reg_r, ldg_ty, ldg_tx, 2)
stage_value_shared(
v_shared_1, kv_reg_r, c1_ldg_ty, c1_ldg_tx, 2
)
T.lma_wait()
T.warpgroup_wait(0)
# T.barrier_arrive(bar_vr0_free)
Expand Down Expand Up @@ -728,9 +740,9 @@ def dsa_decode_split(
elif tid >= 512:
kperm_mask_local = T.alloc_local([4], "bool")
kperm_indices_local = T.alloc_local([4], "int32")
# producer: 128 ldg_ty 16
ldg_tx = (tid - 512) % 8
ldg_ty = (tid - 512) // 8
# producer: 128 threads, 16 producer_ldg_ty rows
producer_ldg_tx = (tid - 512) % 8
producer_ldg_ty = (tid - 512) // 8
ldg_scale_tx = (tid - 512) % 2
ldg_scale_ty = (tid - 512) // 2
for i_i in range(start_block_idx, end_block_idx):
Expand All @@ -746,8 +758,8 @@ def dsa_decode_split(
i_i,
effective_topk,
seq_len_kv,
ldg_ty,
ldg_tx,
producer_ldg_ty,
producer_ldg_tx,
phase_count[0],
kperm_indices_local,
kperm_mask_local,
Expand Down Expand Up @@ -775,10 +787,11 @@ def dsa_decode_split(
k_pe[
kperm_indices_local[r],
g_i,
32 * u + ldg_tx * 4 + v,
32 * u + producer_ldg_tx * 4 + v,
],
kv_shared_l[
r * 16 + ldg_ty, 64 * u + ldg_tx * 8 + v
r * 16 + producer_ldg_ty,
64 * u + producer_ldg_tx * 8 + v,
],
force_async_copy=True,
src_robust_desc=kv_robust_desc,
Expand Down Expand Up @@ -824,9 +837,12 @@ def dsa_decode_split(
k_pe[
kperm_indices_local[r],
g_i,
dim_qk // 2 + 8 + ldg_tx * 8 + v,
dim_qk // 2 + 8 + producer_ldg_tx * 8 + v,
],
k_tail_shared[
r * 16 + producer_ldg_ty,
producer_ldg_tx * 8 + v,
],
k_tail_shared[r * 16 + ldg_ty, ldg_tx * 8 + v],
force_async_copy=True,
src_robust_desc=kv_robust_desc,
)
Expand Down Expand Up @@ -855,10 +871,14 @@ def dsa_decode_split(
k_pe[
kperm_indices_local[r],
g_i,
dim_qk // 4 + 32 * u + ldg_tx * 4 + v,
dim_qk // 4
+ 32 * u
+ producer_ldg_tx * 4
+ v,
],
kv_shared_r[
r * 16 + ldg_ty, 64 * u + ldg_tx * 8 + v
r * 16 + producer_ldg_ty,
64 * u + producer_ldg_tx * 8 + v,
],
force_async_copy=True,
src_robust_desc=kv_robust_desc,
Expand Down