The kernel's `_max_seq_prefix_len` computation unconditionally applied a
causal upper bound on the KV-tile loop:
_max_seq_prefix_len = context_len
+ q_block_local_idx * kBlockQ_dyn
+ (kBlockQ_dyn - 1) + 1
Under causal masking this is the correct optimisation — a Q-block whose
largest row index is R only needs to read K[0..R] because rows beyond R
are softmax-masked to zero. Under `mask_type=0` (no mask) every Q row
must attend to all K rows, so this truncation is incorrect: every
Q-block other than the last one ends up reading a too-short prefix of K
and the resulting softmax / weighted-sum is over the wrong support.
Symptoms at sq=sk=512, hq=hk=5, d=128, bf16, no-mask:
Q-block 0 (rows 0..255): max diff vs fp32 attention_ref ≈ 0.25
Q-block 1 (rows 256..511): max diff vs fp32 attention_ref ≈ 1e-3 (ULP)
The bug never showed up in the cross-impl sweeps because Triton-UA
asserts causal=True (its only supported mode) and sweep_fp8.sh forwards
that default through.
Fix: gate the truncation behind kHasMask. When kHasMask == false the
loop bound is simply `seq_len`, matching the math.
Validated against `aiter.test_mha_common.attention_ref` across:
- MHA d={64,128} sq=sk∈{256..2048} bf16/fp16 no-mask & causal
- GQA-8 d=128 sq=sk∈{256..1024} bf16 no-mask & causal
22/22 stages PASS within bf16/fp16 ULP. sweep_fp8.sh (causal) timings
unchanged — the truncation still fires for the causal kernels.
Co-authored-by: Cursor <cursoragent@cursor.com>
Promote the runtime `page_size` argument to a non-type template parameter
`kPageSize_` on UnifiedAttentionPipeline. Thread it through
unified_attention_kernel_traits and dispatch_variant<V> so the host-side
dispatcher routes on args.page_blk_size ∈ {16, 32, 64} to a constexpr-
pinned prefill instance; values outside that menu (or any decode variant)
fall back to the existing kPageSize_=0 runtime-page-size instance.
Two wins fold together on the prefill tiers:
1. Strength-reduction. Every `/ page_size`, `* page_size`, and `% page_size`
in the per-tile address chain collapses to a literal-folded shift /
multiply-by-magic (`/ 32` → shr 5, etc).
2. Wider Tier-0/Tier-2 gate. The scalar-promote + LDS-cache fast path now
uses the *real* precondition `KY0_step_N <= kPageSize` at compile time
instead of the conservative `KY0_step_N <= 16` hedge — so prefill_d128
bf16/fp16 (KY0_step_N=32), prefill_d64 fp8 (KY0_step_N=32), and
prefill_d64 bf16/fp16 (KY0_step_N=64) also enter the fast path at
their natural page sizes.
Measured impact (sq=sk=75600, MI355, n=30 iters, GQA-8):
variant KY0_step_N ps before after Δ
prefill_d128 fp8 16 32 119.0 111.5 -6.3 %
prefill_d128 bf16 32 32 132.7 130.3 -1.8 %
prefill_d64 fp8 32 32 80.9 68.1 -15.8 %
prefill_d64 bf16 64 64 74.4 73.4 -1.3 %
Decode variants stay on the kPageSize_=0 instances (Tier-0 gate gates them
out anyway — <8 warps — and the binary-size cost isn't justified). All
sweep_fp8.sh shapes + 21 multi-seed multi-sk-length prefill shapes
correctness-PASS. Pre-existing Tier-2 LDS-cache limit (4096 entries)
documented in the pipeline header — same constraint applies to the
kPageSize_=0 fallback so this is not a regression.
36 new prefill instance files: prefill_d{64,128} × {fp16, bf16, fp8} ×
{mask, nmask} × {ps16, ps32, ps64}.
Co-authored-by: Cursor <cursoragent@cursor.com>
GetAlignmentK / GetAlignmentV previously returned a blanket 4 B/lane
(one dword) for every FP8/BF8 tile, citing the gfx950 LDS-direct load
constraint (only dword / dwordx3 / dwordx4 are supported). That cap was
correct for the 8-warp prefill variants (kBlockSize=512, NumIssues drops
to 0.5 at 16 B/lane) but over-applied to every decode tier, where the
1/2/4-warp tile geometry has plenty of headroom.
Refactor the alignment selector into GetKVAlignmentBytes<>, which picks
dwordx4 whenever NumIssues = kPageBlockSize*kHeadDim/(kBlockSize*16)
is an integer >= 1 and falls back to dword otherwise. BF16/FP16 paths
stay at 16 B/lane on every compiled tile, so existing perf is unchanged.
FP8 prefill_d{64,128} also keep the historical dword path because
NumIssues = 0.5 there. FP8 decode_d{64,128}_m{16,32,64,128} now use
dwordx4: same byte volume per K/V tile but 4x fewer async-load issues
(SQ_INSTS_VMEM 131M -> 33M on b=128 sq=1 sk=128000 d=64).
Wall-clock impact on the long-context decode sweep (HIP_VISIBLE_DEVICES=2,
ITERS=20, WARMUP=5, MI355):
shape dtype before after speedup
decode d=64 sq=1 sk=128000 b=128 fp8 7.17 ms 4.57 ms 1.57x
decode d=64 sq=1 sk=128000 b=256 fp8 16.24 ms 9.51 ms 1.71x
decode d=128 sq=1 sk=128000 b=128 fp8 13.11 ms 7.15 ms 1.83x
decode d=128 sq=1 sk=128000 b=256 fp8 31.37 ms 9.78 ms 3.21x
decode d=64 sq=1 sk=128000 b=4 fp8 0.42 ms 0.22 ms 1.92x
decode d=128 sq=1 sk=128000 b=4 fp8 0.80 ms 0.42 ms 1.93x
prefill d=64 sq=75600 sk=75600 b=1 fp8 81.4 ms 81.2 ms 1.00x (dword fallback)
prefill d=128 sq=75600 sk=75600 b=1 fp8 143.5 ms 143.6 ms 1.00x (dword fallback)
Correctness verified across fp8/bf16/fp16, causal/non-causal, and all 7
compiled tile variants. Full PMC + PC-sample analysis is in
ua-test-scripts/rocprof_analysis/BOTTLENECK_ANALYSIS.md section 8.
Co-authored-by: Cursor <cursoragent@cursor.com>
`ADD_SBARRIER_FOR_PHASE0=1` added an extra `s_barrier()` at the start of
every `cl_p` half of every KV iteration, on top of the three barriers
that already gate the LDS hand-offs in phases 1/2/3.
rocprofv3 bottleneck analysis (b=4 sq=8 sk=4096 hq=64 hk=8 d=128 fp8):
the prefill_d128 8-warp variant spends ~15% of GUI_ACTIVE cycles at
s_barriers and shows %any_wait ≈ 200%. PC sampling pinpoints the
phase-0 `s_barrier` (right after softmax rescale, before async prefetch)
as a top hotspot.
Examining the data flow shows the phase-0 barrier is redundant:
- phase1's `s_waitcnt vmcnt(...); s_barrier` guards the K-LDS write
(from the previous iter's K async load) before any warp reads it.
- phase2's `s_waitcnt lgkmcnt(0); s_barrier` guards the softmax-P
LDS write before gemm1 reads it.
- phase3's `s_waitcnt vmcnt(...); s_barrier` guards the V-LDS write
before the next iter's V-LDS read.
These three already provide every cross-warp ordering the pipeline
needs. The phase-0 barrier was purely defensive.
Measurement: 0.1945 → 0.1883 ms (n=300 iters × 3 trials, single shape).
Correctness verified against the Triton reference on fp8/bf16/fp16 ×
{b=4/32/128} × {sq=1/4/8} × {causal,non-causal} × d∈{64,128}.
Leaving the macro and the `=1` documented path in place so the previous
behaviour can be restored if a future arch/shape regresses.
Co-authored-by: Cursor <cursoragent@cursor.com>
Previously the 32x32x16 FP8 P-tile cvt and the QK-C -> PV-A cross-lane
swap ran in two separate static_for loops back-to-back inside fmha_alu1:
the whole tile was cvt'd into p.thread_buf_ first, then a second pass
issued one ds_bpermute_b32 per 8-fp8 K-chunk and read/wrote the same
buffer to swap the "bad" 4-byte halves between paired lanes.
The ds_bpermute has nontrivial LDS-DMA latency that the scheduler has
no way to hide when it lives alone in a tight serial loop with the
gather/scatter packs around it.
Fuse the two into one 8-fp8-per-iter loop:
1. cvt 8 fp32 -> 2 packed uint32 (lo_pack=slot[0..3], hi_pack=slot[4..7])
using the chained cvt_pk_fp8_f32 pattern matching cast_tile_pk_fp8_fp32.
2. Pick own_bad = (sub==0 ? hi_pack : lo_pack) and issue ds_bpermute on it.
3. Write back all 8 fp8 bytes; the "good" half lands first so its byte
stores can overlap with the in-flight ds_bpermute, and the next
iter's cvts can begin while the swap is still pending.
The 16x16x32 LDS-roundtrip branch keeps the original separated cvt
loop (no swap latency to hide there since the relayout goes through
LDS, not ds_bpermute).
Single-shape FP8 perf on gfx950 GPU 2 (CUDA graph, 50 iters):
decode d=128 b=4 sq=8 sk=4096: 0.2106 -> 0.1951 ms (-7.4%)
decode d=64 b=4 sq=8 sk=4096: 0.1464 -> 0.1208 ms (-17.5%)
prefill d=128 b=2 sq=512 sk=4k: 0.2558 -> 0.2220 ms (-13.2%)
BF16 unchanged (0.2046 -> 0.2039 ms, within noise).
Correctness: pytest UA correctness suite 405 passed / 80 skipped
(245 BF16/FP16 + 160 FP8), unchanged from before.
Co-authored-by: Cursor <cursoragent@cursor.com>
The 32x32x16 tiers (prefill_d{64,128}, decode_d{64,128}_m{32,64,128}) keep
the cheap in-register `ds_bpermute_b32` cross-lane swap that fixes the
QK-C / PV-A per-thread alias for the union'd `sp_compute` / `p`.
The 16x16x32 m16 tiers (decode_d{64,128}_m16) cannot use the swap -- the
MFMA puts the paired-lane bit at a different position and the
sub=0/sub=1 4-fp8 chunks no longer map onto each other. We add a
layout-agnostic LDS roundtrip as the `else` branch, gated by the same
`PVWarpTile` constexpr:
- Hoist two distribution-bound windows over the existing `p_lds`
region (one bound to the QK-C output distribution, one to the PV-A
input distribution). Done once per kernel invocation.
- In `fmha_alu1`, after the cvt_pk_fp8_f32 packing chain, view the
union's bytes as a `static_distributed_tensor<fp8>` in the QK-C
distribution, `store_tile` it through `p_lds` in canonical (M, N)
order, `s_barrier`, then `load_tile` back with the PV-A
distribution and copy into `sp(idx).p`.
A/B'd a uniform LDS-roundtrip (no fast-path) vs the split: pure LDS
regressed decode_m128 by ~1.5x end-to-end (CK FP8 dropped from
~0.39x of Triton FP8 to ~0.16x), driven by the extra block-wide
barrier on the 4-warp decode path. Keeping the swap for 32x32x16
preserves the previously-tuned perf.
Dispatcher (`unified_attention.cpp`) now FP8-enables every UA variant
including decode_d{64,128}_m16. Four new instance .cpp files
(`unified_attention_d{64,128}_fp8_{mask,nmask}_decode_t.cpp`)
instantiate the m16 FP8 kernels.
Pytest (`test_unified_attention_ck_correctness.py`):
- 245 BF16/FP16: pass (no regression from the pipeline edit).
- 160 FP8: pass (was 112 before m16 enablement).
- 80 skipped: block_size<32 or query_len>kv_len -- pre-existing.
Single-shape m16 dispatches verified on gfx950:
b=128 sq=1 hq=hk=8 d=128 fp8 PASS (CK 0.109 ms / Triton 0.043 ms)
b=128 sq=1 hq=hk=8 d=64 fp8 PASS (CK 0.077 ms / Triton 0.039 ms)
Co-authored-by: Cursor <cursoragent@cursor.com>
Full pipeline support for FP8 (e4m3fn on gfx950 / e4m3fnuz on gfx942)
in the unified-attention kernel, gated to the 32x32x16 MFMA tiers in
both d=64 and d=128 ladders: prefill_d{64,128}, decode_d{64,128}_m128,
decode_d128_m32, and decode_d64_m64. The 16x16x32 _m16 tiers stay
BF16/FP16-only -- the QK-C and PV-A per-thread layouts there differ
by an M<->N swap that the current slot-swap fixup cannot express; a
full per-thread transpose (most likely via LDS) is needed.
Pipeline (unified_attention_pipeline.hpp):
* `fmha_alu1` now performs a cross-lane P-tile fixup right after the
FP8 packing of softmax(P). It's a `ds_bpermute_b32` between paired
lanes `lane ^ 32`, swapping sub=0 slot[k_base+4..k_base+7] with
sub=1 slot[k_base..k_base+3] for every 8-fp8 chunk. This realigns
the FP8 packed P operand with PV-A's `Single` AttrNumAccess
per-thread layout, which is necessary because the QK-C output and
PV-A input alias byte-for-byte via the sp_compute/p union -- and
for FP8 the two warp-gemm layouts no longer agree (BF16/FP16 keep
Double AttrNumAccess in the PV gemm, which matches QK-C natively).
Gated on `Gemm1WarpTile == 32x32x16`; FP8-only (BF16/FP16 paths take
the existing cvt_pk path unchanged).
Default policy (unified_attention_pipeline_default_policy.hpp):
* PV warp gemm now selects `WGAttrNumAccessEnum::Single` when V is
fp8/bf8 and `Double` otherwise. Forced by load_tile_transpose's
SubMinDim = 64-bit / sizeof(V) constraint: for FP8 SubMinDim=8 and
kABKPerLane=8 only Single satisfies the validation static_asserts.
* GetAlignmentK / GetAlignmentV on gfx950 drop to 4 B/lane for fp8/
bf8. The natural 16 B/lane async-load that BF16/FP16 use leaves
NumIssues = 0 for the FP8 tile shapes we compile, and 8 B/lane
fails the dword / dwordx3 / dwordx4 constraint in
amd_buffer_addressing_builtins. 4 B/lane gives NumIssues >= 1 on
every targeted variant and is the same alignment the gfx942
fallback already used. BF16/FP16 keep the full 16 B/lane path so
existing perf is unchanged.
* GetSmemSizeKV adds a `VLoadDescSize` lower bound. The
MakeVLdsLoadBlockDescriptor's element span dominates the banked
SingleVSize only for FP8 (small per-lane KVector + fixed
kVLdsPadInBytes = 64), so without it FP8 hits the GetSmemSizeKV
static_asserts. BF16/FP16 are unaffected.
Warp-gemm headers + dispatcher:
* New `WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed_T<AttrNumAccess>`
template alias in warp_gemm.hpp (mirrors the existing BF16 32x32x16
CTransposed template), used by the PV gemm to thread the FP8
Single AttrNumAccess through.
* New Dispatcher specialization for
<fp8_t, fp8_t, float, 32, 32, 16, true, false, false, EDouble>
in warp_gemm_dispatcher.hpp routing to the new template.
ABI / dispatcher (unified_attention.{cpp,hpp}, unified_attention_impl.hpp):
* New `fp8` value in `unified_attention_args::data_type_enum` (selects
e4m3fn on gfx950 via CK_TILE_USE_OCP_FP8, e4m3fnuz elsewhere).
* New `unified_attention_problem_traits<...::fp8>` alias:
qkvp_dtype = ck_tile::fp8_t, acc_dtype = float, o_dtype = bf16_t
(matches the Triton reference), lse_dtype = float.
* Per-tensor `q_descale` / `k_descale` / `v_descale` floats on
`unified_attention_args` (default 1.0f so non-FP8 round-trips
cleanly). The pipeline folds q_descale*k_descale into the softmax
scale and applies v_descale once to o_acc after the 1/l norm --
same semantics as Triton's q_scale/k_scale/v_scale.
* `dispatch_variant<>` enables FP8 on prefill_d{64,128},
decode_d{64,128}_m128, decode_d128_m32, decode_d64_m64. The
16x16x32 _m16 tiers return (false, -1.f) for now (see top comment).
Instances:
* 12 new FP8 .cpp files under example/.../42_unified_attention/
instances/ covering the 6 enabled variants x {mask, nmask}.
Validation: 112 / 0 / 128 in the FP8 pytest sweep (passed / failed /
m16-skipped); 245 / 245 in the BF16/FP16 sweep (no regression).
Functional correctness is within the FP8 quant-noise tolerance the
Triton FP8 suite uses (atol/rtol = 1.5e-1). Perf still trails Triton
across the enabled tiers (CK FP8 / Triton FP8 = 0.39-0.69x on the
shapes we benchmarked); that's a separate workstream.
Co-authored-by: Cursor <cursoragent@cursor.com>
Replace the speculative TODO-style comment next to the K_mem_load /
V_mem_load dispatch with a record of the actual experiment: we
implemented async_load_tile_raw_rebased (buffer_load_dword_lds with a
per-issue SRD whose 48-bit base absorbs the wave-uniform page offset),
verified correctness on multiple big-cache decode shapes, and measured
it against the existing async_load_tile_raw_long path on an isolated
GPU. Rebased was at best tied with long and at worst ~6% slower
(b=1 sk=1M d=64 GQA8: 2.46 ms vs 2.32 ms; b=8 sk=200k d=128 GQA8:
2.12 ms vs 2.02 ms). The workloads are compute / softmax bound, not
K/V load bandwidth bound, so the buffer_load throughput edge never
materialises, while the per-issue SRD construction adds real SGPR
pressure.
No functional change in this commit -- only the explanatory comment is
updated so the next person who eyes the same idea finds the receipts
before re-implementing.
Co-authored-by: Cursor <cursoragent@cursor.com>
The shared-SRD buffer_load_dword_lds path that K_mem_load / V_mem_load use
wraps the per-lane voffset (int32 bytes) once
num_blocks * page_size * row_stride * sizeof(T) > INT32_MAX,
silently returning wrong data on large paged-KV pools (e.g. >4 GB caches).
Add a second path, async_load_tile_raw_long, that issues the same load via
__builtin_amdgcn_global_load_lds with per-lane 64-bit base pointers, lifting
both 4 GB limits (SRD size + voffset). Per-issue LDS pointers are computed
explicitly because the intrinsic sets m0 itself, so the old m0_set / m0_inc
bookkeeping doesn't apply. The path also clamps lane_elem_off to the live
buffer range to mimic the original SRD's hardware OOB behaviour.
Dispatch is a wave-uniform runtime branch on a new
cache_ptr_int32_overflow_possible flag plumbed from
unified_attention_args through MakeKargs into the pipeline operator().
Small caches keep the original buffer_load throughput; only the (rare)
>4 GB cache pays the global_load_lds cost.
k_page_offsets / v_page_offsets are widened to long_index_t. The original
buffer_load path implicitly narrows back to int32 when forwarding through
async_get_vectorized_elements_raw, which is intentional and safe whenever
the overflow flag is false.
For diagnostics, also derive a constexpr KWaveSpanInN =
(LaneGroups - 1) * NumWarps + 1 inside the pipeline; when this exceeds
page_size a single buffer_load spans multiple random pages, so the
per-issue SRD-rebase optimisation (not implemented yet) would not apply
even on a sub-4 GB cache. Informational only today.
Test: ua-test-scripts correctness sweep (245/245 pass), plus
test_single_shape.py -b 32 -sq 8192 -sk 120000 -hq 64 -hk 8 -d 64 \
--num-blocks 1200000 --block-size 16 --test
which previously returned wrong data due to the int32 wrap and now passes
with max abs diff 1.22e-04 vs Triton.
Co-authored-by: Cursor <cursoragent@cursor.com>
After moving kBlockQ to runtime in the previous commit, the static
NumQPerKV in `variant_config<V>` and the runtime-vs-static assert in
the kernel became the only things still tying a compiled binary to a
specific num_queries_per_kv. Drop both and the existing instances now
serve every num_qpkv that divides kBlockM evenly.
Concretely:
* `variant_config<V>` -- remove the NumQPerKV field from every
specialization.
* `unified_attention_kernel_traits` -- remove the `num_queries_per_kv`
/ `kBlockQ = kBlockM / num_qpkv` derivation. The BlockTile's 2nd
entry (the static `kBlockQ` exposed via UnifiedAttentionShape) is
anchored at kBlockM so it describes the "num_qpkv == 1" fallback;
the actual kBlockQ is always the runtime value.
* `unified_attention_kernel_launch` -- recompute kBlockQ at host time
from `args.num_queries_per_kv` for the total_num_q_blocks math.
* `unified_attention_kernel.hpp` -- drop the
`assert(kBlockQ_dyn == kBlockQ)` (it enforced the very coupling we
just removed).
* `unified_attention.cpp::select_config` -- collapse the two
per-num_qpkv code paths into a single (head_dim, avg_rows,
max_rows) ladder, where avg_rows = avg_q * num_qpkv.
Variant renames (8 variants):
prefill_d128_mha -> prefill_d128
decode_d128_mha_m128 -> decode_d128_m128
decode_d128_mha_m32 -> decode_d128_m32
decode_d128_mha_m16 -> decode_d128_m16
prefill_d64_gqa8 -> prefill_d64
decode_d64_gqa8_m128 -> decode_d64_m128
decode_d64_gqa8_m64 -> decode_d64_m64
decode_d64_gqa8_m16 -> decode_d64_m16
The 16 d=64 instance files lose their `_gqa8` infix to match the
d=128 naming (file count unchanged: 16 dtypes x mask combos per
head_dim).
Validation:
* Correctness suite: 241/245 (same 4 pre-existing int32-overflow
failures in the prefill rebased-pointer path).
* d=128 GQA-8 (a NEW combo we never had a binary for) -- runs
correctly on the existing decode_d128_m* binaries with num_qpkv=8
at runtime. max abs diff <= 1e-2 vs the torch reference at ql in
{1, 4, 16}.
* d=64 MHA (also a new combo) -- runs correctly on the existing
decode_d64_m* binaries with num_qpkv=1. Same tolerance.
* Perf sweep (b=4..256, sk=120000, MI300):
d=64 GQA-8: speedups 1.28x..1.84x vs Triton (within 0.6%
of baseline).
d=128 MHA: speedups 0.98x..1.14x vs Triton (within 0.3%
of baseline).
Unlocked: adding new (head_dim, num_qpkv) combos no longer requires
new kernel binaries -- just a host-side heuristic update mapping the
combo to the appropriate (kBlockM, BlockWarps) ladder.
Co-authored-by: Cursor <cursoragent@cursor.com>
kBlockQ (= kBlockM / num_queries_per_kv) was constexpr in
`UnifiedAttentionShape` / the kernel-traits, forcing one kernel
instance per (kBlockM, num_qpkv) pair even though the matmul tile is
fully determined by kBlockM and kHeadDim. Audit confirmed kBlockQ
only feeds:
* arithmetic in `unified_attention_kernel.hpp` (loop bounds, Q-tile
indexing, query_len padding),
* `pad_tensor_view` size tuples for Q/O/LSE DRAM views,
* one `mask.IsEdgeTile(... number<kBlockQ>{} ...)` call inside the
pipeline's per-K-tile mask check.
None of these structurally need a compile-time value:
* `pad_tensor_view` already accepts mixed runtime/compile-time tuple
elements (e.g. it's passed plain `1` next to `kHeadDimPadded`).
* `IsEdgeTile` only does runtime arithmetic on the tile size; adding a
runtime overload that accepts `index_t` is trivial (the compile-time
one now forwards to it).
Wiring:
* `block_masking.hpp` -- add an `IsEdgeTile(..., index_t tile_h,
index_t tile_w)` overload; the existing `number<>` overload just
forwards to it.
* `unified_attention_pipeline.hpp` -- new optional
`num_queries_per_kv` arg on the pipeline's `operator()` (default 0
keeps existing call sites unchanged). Computes
`kBlockQ_dyn = (num_qpkv > 0) ? (kBlockM / num_qpkv) : kBlockQ`
once at the top, uses it in the IsEdgeTile call.
* `unified_attention_kernel.hpp` -- compute
`const index_t kBlockQ_dyn = kBlockM / kargs.num_queries_per_kv`
once and replace every per-call `kBlockQ` use with `kBlockQ_dyn`.
Pass `kargs.num_queries_per_kv` through to the pipeline. The
debug-only assert(`kBlockQ_dyn == kBlockQ`) keeps the static and
dynamic values in lock-step until we actually collapse variants.
Perf A/B (b=4..256, sk=120000, MI300):
d=128 MHA (num_qpkv = 1, runtime div is trivial):
BW within +/-0.2% across all batch sizes (noise).
d=64 GQA-8 (num_qpkv = 8, runtime division actually happens):
speedups 1.28x..2.14x vs Triton -- identical to baseline.
Correctness suite stays at 241/245 (same 4 pre-existing int32-overflow
failures in the d=128 prefill rebased-pointer path).
This is a no-op on perf and unlocks a follow-up where we collapse the
two num_qpkv values per (head_dim, kBlockM) -- e.g. the future d=128
GQA-8 variant can reuse the existing decode_d128_mha_* instances by
just passing a different runtime num_queries_per_kv.
Co-authored-by: Cursor <cursoragent@cursor.com>
End-to-end split-KV (FlashDecoding-style) for the CK unified attention
kernel. The host launches a single 3D grid with z == num_splits; each
CTA computes its KV-range slice and writes a normalized (o_acc, lse)
partial to FP32 workspaces, which the caller reduces into the final
output.
Pipeline changes:
- operator() returns ck_tile::make_tuple(o_acc, lse) instead of just
o_acc. The masked-empty early-exit returns lse = -inf so downstream
combine weighs the partial as zero.
- LSE is built in the natural-log domain from the pipeline's *unscaled*
rowmax: lse = (scale_s / log2(e)) * m + log(l). Previously we used
m / log2(e) + log(l), which dropped the per-head scale and produced
LSE values ~1/scale too large.
- Fix post-process parity: which SP register is left in the
alu0-done/alu1-pending state at loop exit depends on the parity of
the *iteration count* (= num_total_loop - num_blocks_start), not on
num_total_loop alone. For non-split (num_blocks_start == 0) the two
parities coincide; for splits starting at an odd tile they don't.
- Fix split-KV page-table offset: num_blocks_start is counted in
kPageBlockSize-sized tiles, but block_tables is indexed in
page_size-sized pages — shifting block_table_offset by num_blocks_start
reads the wrong pages whenever kPageBlockSize != page_size. Replaced
with split_token_offset = num_blocks_start * kPageBlockSize added to
logical_token before /page_size, so the page lookup uses the absolute
token position.
Kernel + dispatcher:
- Drop kargs.i_split; each CTA reads i_split = blockIdx.z.
- GridSize{2D,Decode} now take num_splits and add it as the z-dim
(defaults to 1, so non-split callers see dim3(..., 1, 1)).
- New write path: when num_splits > 1, the kernel skips the user
epilogue and instead writes the FP32 (o_acc, lse) tile pair into
workspace tensors at [head, split, batch_start_token, ...] using
Default2DEpilogue (UseRawStore=true) for o_acc and store_tile for
lse. Host strides come from kargs.
Co-authored-by: Cursor <cursoragent@cursor.com>
Refactor the K/V DRAM access in the unified-attention pipeline to use
tile_scatter_gather with a unified per-(thread, Y0-iter) page-offset
formula:
logical_token = tile_idx * kPageBlockSize + thread_N_pos + i * Y0_step_N
logical_page = logical_token / page_size
within_page = logical_token % page_size
phys_page = block_tables[block_table_offset + logical_page]
page_offsets[i] = (phys_page * page_size + within_page) * row_stride
The page indirection now lives entirely in page_offsets, refreshed via
update_page_idx() between iters. The per-iter SRD rebase
(set_bottom_tensor_view_data_ptr + init_raw) and the use_ptr_rebase
overflow heuristic are gone.
Effects:
- The assertion kv_page_size_in_blocks >= 1 (i.e. kPageBlockSize <=
page_size) in the kernel is dropped. Tiles may now span multiple
cache pages, as long as Y0_step_N (= N1*N2 from the K/V tile dist)
divides page_size so that a wave-wide load never straddles a page.
- Pipeline arg renamed kv_page_size_in_blocks -> page_size (PageSize
in tokens). Kernel passes kargs.page_size through directly.
- Validated correctness vs Triton on bf16 / d=64 / decode_s with
block_size in {16, 32, 64}; max abs diff 1.22e-04 in all cases.
Perf is on par with the prior pass-1 scaffolding (~3.6 ms on the
131072-context shape).
TODO(overflow): page_offsets are index_t; caches whose
num_blocks * page_size * row_stride exceeds INT32_MAX will wrap.
A future change should plumb long_index_t through the scatter-gather
load path or compute a per-batch min-page shift in a pre-pass.
TODO(unsupported regime): page_size < Y0_step_N (a wave crosses a page
mid-iter) needs per-lane VGPR SRDs and is not implemented.
Co-authored-by: Cursor <cursoragent@cursor.com>
tensor_coordinate::get_offset() returns index_t (int32), causing overflow
when page_idx * block_size * stride > 2^31 (~131K blocks for d64/GQA-8).
Fix: rebase K/V data pointer for each page using int64 arithmetic instead
of set_window_origin with large offsets. After rebasing p_data_ and
buffer_size_, call init_raw() to refresh the AMD buffer resource descriptor,
then set_window_origin({0,0}) to reset cached coordinates.
Tested: num_blocks up to 2M with nkh=1/8, blk=32/64. All pass.
Made-with: Cursor
The final V tile's async load was not properly waited on before reading
from LDS: s_waitcnt_vmcnt<K_inst> allowed V_inst outstanding loads
(a no-op when K_inst == V_inst). The last loop iteration never prefetches
K, so only V is outstanding. Use s_waitcnt_vmcnt<0> unconditionally.
This partially fixes the BS32 race condition for production workloads
(maxk >= 256). A deeper pipeline race remains for very short KV
sequences (maxk < ~165, 2-5 pages) with block_size=32 at high batch.
Made-with: Cursor
Added split-KV fields to UnifiedAttentionVarlenKargs (num_splits,
i_split, lse_acc_ptr, o_acc_ptr + strides). Modified operator() to
compute per-split KV range using blocks_per_split.
INCOMPLETE: The pipeline returns normalized o_acc but the split-KV
combine kernel needs unnormalized o_acc + lse. Need to modify the
pipeline to optionally return m and l values alongside o_acc.
The kernel changes compile but the epilogue needs the split path
(write to float accumulators instead of final output).
Made-with: Cursor
The kSkipMinSeqlenQ optimization incorrectly used <= comparison, causing
the kernel to skip batches where seqlen_q equals min_seqlen_q. This
happens in the common case of no padding (all batches have the same
seqlen_q == min_seqlen_q), producing all-zero output silently.
Changed to strict < so batches with exactly min_seqlen_q tokens are
still processed.
Made-with: Cursor
[CK] Fix async pivot mismatch in persistent GEMM kernel
scheduler (#5776)
## Motivation
Fix pivot mismatch in the persistent GEMM kernel's async input scheduler
that causes **GPU hangs** and incorrect results when used with AsyncTP
(Asynchronous Tensor Parallelism) on ROCm.
PyTorch's `_fused_all_gather_matmul_native` uses this persistent GEMM
kernel with chunk signals to overlap communication and computation. The
pivot mechanism ensures each rank starts computing from its own local
shard first (which is already available), then moves to remote chunks as
they arrive over the network.
Because of the pivot mismatch, the kernel frequently waits on signals
for chunks that have not yet arrived, while attempting to read data from
completely different chunks. This synchronization desync reliably
triggers infinite hangs during multi-GPU native AsyncTP execution. This
fix is required to enable functional AsyncTP support on ROCm.
## Technical Details
In the persistent kernel loop (`UniversalGemmKernel::operator()`), the
M-tile coordinate used for data selection (`i_m`) and the M-tile
coordinate used for the chunk-signal wait (`chunk_idx`) were derived
from inconsistent bases:
* `i_m` was computed from the **unpivoted** tile index `iM`.
* `chunk_idx` was computed from the **pivoted** expression `(iM +
tile_idx_pivot)`.
This means the kernel could wait for chunk N's signal but then read from
chunk M's memory, or vice versa. The mismatch scales with GPU count:
with 2 GPUs ~50% of tiles are wrong, with 4 GPUs ~75%, etc.
**The Fix:**
Introduce a single pivoted M-tile index (`iM_eff`) and derive both `i_m`
and `chunk_idx` from it. This guarantees the kernel always waits for the
correct chunk before reading its data.
*(Note: Minor cosmetic `clang-format` changes were also pulled in
alongside the fix).*
## Test Plan
1. Build PyTorch with this CK change.
2. Run the specific multi-GPU AsyncTP native test:
`timeout 180s env HIP_VISIBLE_DEVICES=0,1 pytest
test/distributed/test_symmetric_memory.py -k
test_fused_all_gather_matmul_native -q -s -x`
## Test Result
Tests verify correct overlapping execution without hangs or accuracy
mismatches when running the AsyncTP native path with non-zero pivots.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Add pooling in tile_engine
## Motivation
<!-- Explain the purpose of this PR and the goals it aims to achieve.
-->
Add pooling in ck tile engine
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Fix FMHA BWD IGLP incorrect results due to AGPR
misallocation (#5991)
## Motivation
After PR #5790 removed the `if constexpr(FmhaMask::IsMasking)` guard
around the
`num_total_loop <= 0` early-exit check, the IGLP pipeline
(`BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP`) produces incorrect dK/dV
gradients for
non-masking kernels (even with fix in #5915). Assembly inspection
confirms that the CFG change causes the LLVM
register allocator to reuse AGPR accumulators as scratch destinations in
the dK/dV
reduction loop, breaking the loop-carried accumulation across Q-tile
iterations.
## Technical Details
- Add `[[unlikely]]` to the `num_total_loop <= 0` early-exit in
`BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP`. This attribute is load-bearing:
it
restores the CFG shape that the register allocator needs to correctly
assign
dedicated AGPRs to each column of the dK/dV accumulator.
- Only the IGLP pipeline is affected; the other two BWD pipelines do not
exhibit
this issue.
## Test Plan
## Test Result
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] contraction: extend GetTypeString() to include
layout-differentiating params (#6022)
## Motivation
Consumers that identify kernels by their `GetTypeString()` (such as
hipTensor's actor-critic kernel selection, which hashes the string into
a
stable cross-platform UID) were silently dropping one of two colliding
variants during registry insertion.
`GetTypeString()` in `DeviceContractionMultipleD_Xdl_CShuffle`
previously
printed 13 template parameters, omitting
`ABlockTransferSrcScalarPerVector`,
`BBlockTransferSrcScalarPerVector`, `ABlockLdsExtraM`, and
`BBlockLdsExtraN`.
These four parameters determine the block-transfer access width and LDS
padding strategy, and are precisely what differentiates the `kk`, `kn`,
`mk`, and `mn` layout variants from one another when all other geometry
parameters are equal. Two instantiations with identical 13-parameter
strings
are distinct C++ types that accept different stride layouts and reject
each
other's arguments via `IsSupportedArgument`.
This patch extends the output to 17 parameters so that every distinct
template instantiation of this class produces a unique
`GetTypeString()`.
## Technical Details
`include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp`:
- extend `GetTypeString()` from 13 to 17 parameters including
`ABlockTransferSrcScalarPerVector`,
`BBlockTransferSrcScalarPerVector`, `ABlockLdsExtraM`, and
`BBlockLdsExtraN`.
## Test Plan
Build CK and hipTensor with these changes, and verify hipTensor can
differentiate and select the
correct kernels with layout variations.
## Test Result
CK is building correctly and hipTensor is selecting the kernels
correctly.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK][CK Tile] Force padding for atomic_add bf16 C tensor
(#5842)
## Motivation
Force padding for atomic_add bf16 C tensor to avoid memfaults.
## Technical Details
- add global atomic add for bf16 and enable them
- add padding for atomic add bf16 due to the lack of oob
- remove padding for not continous dims in conv for other cases
- minor bwd data conv fixes
## Test Plan
test_grouped_conv_*_tile
## Test Result
pending
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Changed cshuffle LDS descriptor to naive layout
(#5729)
## Motivation
This PR changes gemm/convolution cshuffle layout into plain one. to
improve cshuffle operation performance.
## Technical Details
The purpose is that before this change the cshuffle layout was having
some descriptor transformations that were probably aimed at reducing LDS
bank conflicts, but the transformations itself were terribly slow, which
negatively impacted the performance.
## Test Plan
There is no need for additional tests, since current tests cover this
functionality.
[CK] fix clang lifetimebound errors with staging compiler
(#5921)
## Motivation
The ROCm staging compiler (newer Clang) enforces
`[[clang::lifetimebound]]` annotations on methods that return references
or pointers to internal object data. Without these annotations, the
staging compiler emits compilation errors for container accessor methods
across the CK and CK Tile namespaces.
## Technical Details
Adds `[[clang::lifetimebound]]` to all reference/pointer-returning
accessors in core container types:
**`ck::` namespace:**
- `Array` -- `At()`, `operator[]`, `operator()`, `begin()`, `end()`
- `index_array` -- `operator[]`
- `StaticallyIndexedArray_v2` -- `At()`, `operator[]`, `operator()`
- `IndexLookupTable` -- `operator[]`
**`ck_tile::` namespace:**
- `array` -- `get(i)`, `at()`, `operator[]`, `operator()`
- `static_array` -- `operator[]`
- `thread_buffer` -- `get(i)`, `at()`, `operator[]`, `operator()`
- `make_kernel()` -- parameter pack
Also removes the unused `instance_index` variable from
`batched_gemm_reduce_fp16.cpp` and simplifies its argument parsing
accordingly.
## Test Plan
- Compile with the staging compiler to verify all lifetimebound errors
are resolved
- Existing tests pass unchanged -- the attribute is a compile-time
annotation with no runtime effect
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Fix FMHA BWD register pressure by wrapping
num_total_loop with amd_wave_read_first_lane (#5915)
## Motivation
In three FMHA backward pipelines, `num_total_loop` is computed without
`amd_wave_read_first_lane()`, so the compiler treats it as a VGPR even
though it is logically uniform across all lanes. This raises register
pressure, and under high pressure the compiler may reuse VGPRs across
overlapping live ranges. This was confirmed via assembly inspection: the
compiler reused `v52:v53` as both the B-matrix input for dK MFMAs and an
intermediate value for dV, producing incorrect dK/dV gradients.
## Technical Details
Wrap `num_total_loop` with `amd_wave_read_first_lane()` in three
pipelines:
- `block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr`
- `block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr_iglp`
- `block_fmha_bwd_dq_dk_dv_pipeline_trload_kr_ktr_vr`
This promotes `num_total_loop` to an SGPR, eliminating the excess
register pressure and the incorrect VGPR reuse.
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] More lifetime-warning suppression
## Motivation
The staging compiler picked up another change from upstream that leads
to more lifetime-analysis warnings. This breaks the build, given CK is
built with -Werror. As a result, compiler promotion is blocked.
## Technical Details
This patch adds the pragma push diagnostics to ignore the
lifetime-warnings in the modified files to unblock compiler promotion.
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE ]Revert "[CK_TILE] Enable MXFP6 for MX GEMM op
(#5095)" (#5849)
This reverts commit 7e55766ddf7e9e20791b0e4e2d7b4026cf16b637.
## Motivation
<!-- Explain the purpose of this PR and the goals it aims to achieve.
-->
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Fix min k_batch calculation in conv kernels
## Motivation
Avoid division by 0 and remove not needed "-1".
## Technical Details
Our div up implementation return lower value if input is divisible.
There is no need to subtract 1.
## Test Plan
test_grouped_conv_bwd_weight
## Test Result
Passed locally.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-1019
[CK_TILE] Fix NaN for FMHA BWD When seq_q=0
## Motivation
This PR addresses NaNs in the FMHA backward (dQ/dK/dV) path when the
effective query sequence length for a tile is zero, by ensuring the
per-tile pipelines exit early with zeroed accumulators and by avoiding
an early kernel return that prevented writing out cleared gradients.
## Technical Details
- Add unconditional early-exit in the dK/dV pipelines when
`num_total_loop <= 0` (no work), returning zeroed accumulators.
- Adjust group-mode kernel early-return logic to only return when
**both** `seqlen_q` and `seqlen_k` are zero, allowing blocks to run and
store cleared dK/dV when `seqlen_q == 0`.
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK][CK_TILE] Revert addional oob check in gemm IsSupported
function (#5789)
## Motivation
fix ck_tile's oob check.
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE, CK_BUILDER] Add bwd data to CK Tile profiler
(#5516)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
We want close the performance gap between old CK and CK Tile for bwd
data convolutions. To achieve this, we need tow things
- Configurations for the old CK kernel instances such that we can map
them into CK Tile instances.
- Support in CK profiler to run the CK Tile instance with the same API
as for old CK instances.
## Technical Details
Extracted kernel configurations from old CK. The codegen python script
for CK Tile convs is extended to support also bwd data. The generated
instances are added to the CMake build (target
`device_grouped_conv_bwd_data_tile_instances`).
A new profiler op (`grouped_conv_bwd_data_tile`) has been added to the
CK Profiler. The API is same as for old CK's profiler op
`grouped_conv_bwd_data`.
Revert "Ck/joye/revert oob check (#5640)"
This reverts commit 552ab4880292694cb8261f40fa4223af52cb8419.
## Motivation
<!-- Explain the purpose of this PR and the goals it aims to achieve.
-->
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK][CK Tile] Fix kbatch check in grouped conv and gemm
kernels (#5555)
## Motivation
Fix kbatch check in grouped conv and gemm kernels, allow tails for
kbatch.
## Technical Details
Round up K / Kperxdl and divide it by Kbatch to allow tail for K.
## Test Plan
test_grouped_convnd_bwd_weight_tile
## Test Result
passed locally
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK][CK Tile] Improve access for merged groups and remove
modulo from xor (#5334)
## Motivation
[CK][CK Tile] Improve access for merged groups and remove modulo from
xor
## Technical Details
- add template parameter to xor if modulo is needed. We don't need
modulo for merged groups
- use access by m for merged groups for a tensor
-
## Test Plan
test_grouped_convnd_fwd_tile
## Test Result
passed locally
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Ck/joye/revert oob check
## Motivation
fix ck_tile's oob check.
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Rename Stream-K grid function
## Motivation
This PR introduces a change in the name of the get_grid function in the
Stream-K TilePartitioner to avoid confusion with a similarly named
method. In the Stream-K TilePartitioner, there is get_grid() which
returns num_cu*occupancy and there is grid_size() which returns the grid
size used to launch the kernel. In this PR, we change get_grid() to be
get_max_active_wgs() to better reflect what the function returns and not
confuse it with grid_size().
## Technical Details
Initially in the Stream-K TilePartitioner we had get_grid() which
returned grid_. We are renaming get_grid() to get_max_active_wgs() and
grid_ to max_active_wgs_ internally, while keeping grid_size() the same.
The parameter, grid, for the Stream-K TilePartitioner remains the same
to maintain consistency with the rest of the Stream-K API.
## Test Plan
Validated using the test suite that is already present.
## Test Result
All tests passed
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Enable MXFP6 for MX GEMM op
## Motivation
Add support for MXFP6 in the MX GEMM op in CK-Tile.
Depends on https://github.com/ROCm/rocm-libraries/pull/4594
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] add tf32 support
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Proposed changes
TF32 is added in CK on gfx942 and gfx950. This PR is to initiate tf32 in
CK_TILE on gfx942 and gfx950.
## Checklist
Please put an into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [x] I have run on all changed files
- [ ] Any dependent changes have been merged
## Discussion
[CK] Fix MOE FP8 SplitK buffer descriptor OOB
When SplitK is enabled, kernel entry shifts A/B/AScale/BScale base
pointers by SplitKBatchOffset, but make_dynamic_buffer element spaces
are still based on full K dimension. This causes hardware buffer
resource descriptors to extend beyond the actual tensor allocation,
leading to GPU memory access faults when the tensor happens to be placed
at the end of an allocated memory pool region.
Fix by subtracting the split offset from each buffer's element space in
both Run() (v1 pipeline) and Run_2Lds() (v2/v3 pipeline), so the buffer
descriptor range [shifted_base, shifted_base + reduced_space) exactly
covers the valid allocation.
Also refactor SplitKBatchOffset to accept const Problem& (instead of
Argument&) and add a default constructor, enabling direct reuse in
Run/Run_2Lds without duplicating offset calculation logic.
Made-with: Cursor
## Motivation
<!-- Explain the purpose of this PR and the goals it aims to achieve.
-->
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
CK Tile MX GEMM Packing Improvement
## Motivation
Reduce the scale loading size and also has better utilization of MFMA
scale selection.
## Technical Details
Add up the packing of mx scales.
## Test Plan
Use the existing test cases.
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Add LLC-aware FMHA head grouping and head-major
scheduling on RDNA (#5018)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
Long-sequence FMHA can become memory-bound when K/V working sets exceed
Infinity Cache (LLC), causing repeated DRAM traffic across heads.
This PR introduces LLC-aware launch ordering improvements for FMHA
forward, and it is currently enabled only on gfx11 and gfx12. The
approach is inspired by
[`Dao-AILab/flash-attention#2217`](https://github.com/Dao-AILab/flash-attention/pull/2217),
adapted to CK’s kernel/runner structure and layout handling.
In this context, `bshd` is the layout used in Flash-Attention, while
`bhsd` is the default layout used by the CK Tile FMHA example.
## Technical Details
This PR adds two complementary strategies:
- For `bshd` input layout (`i_perm/o_perm=0`), enable explicit LLC-aware
head grouping:
- Estimate LLC size (env override, KFD sysfs, or arch default).
- Compute group size from K/V bytes per head vs LLC target.
- Launch FMHA forward repeatedly per head-group by slicing Q/K/V/O (and
related tensors).
- For `bhsd` input layout (`i_perm/o_perm=1`), apply implicit
launch-order adjustment:
- Keep a single kernel launch.
- Reinterpret block linearization in `GetTileIndex` to make execution
head-major,
improving temporal locality of per-head K/V reuse.
Additional integration updates:
- Propagate `num_head_q_total` and `head_start` through FMHA args/kargs.
- Use global head indexing for dropout RNG stream mapping so grouped
launches keep
deterministic/consistent dropout behavior.
- Keep fallback behavior unchanged when grouping is not beneficial or
disabled.
## Test Plan
- `test_ck_tile_fmha`
- `tile_example_fmha_fwd`
## Test Result
- `test_ck_tile_fmha`: all tests passed.
- `tile_example_fmha_fwd`: tested this on gfx1100, gfx1151, and gfx1201,
and all of them show higher performance compared to the baseline. The
improvement is consistent, and performance is well maintained even at
long sequence lengths.
./build/bin/tile_example_fmha_fwd -prec=bf16 -mode=0 -b=1 -h=24 -d=128
-s={seqlen} -s_k={seqlen} -lse=0 -iperm={0/1} -operm={0/1}
- TFLOPs by sequence length target: gfx1100 layout: bhsd
SeqLen | Before | After | Speedup
-- | -- | -- | --
1024 | 56.27 | 61.48 | 1.09x
4096 | 67.10 | 72.27 | 1.08x
8192 | 65.99 | 71.64 | 1.09x
12288 | 61.60 | 76.61 | 1.24x
16384 | 58.99 | 75.74 | 1.28x
20480 | 57.32 | 74.42 | 1.30x
24576 | 56.89 | 74.25 | 1.31x
27280 | 18.93 | 24.48 | 1.29x
- TFLOPs by sequence length target: gfx1201 layout: bshd
SeqLen | Before | After | Speedup
-- | -- | -- | --
1024 | 66.79 | 65.90 | 0.99x
4096 | 85.90 | 86.80 | 1.01x
8192 | 77.06 | 90.29 | 1.17x
12288 | 58.36 | 88.98 | 1.52x
16384 | 52.12 | 88.88 | 1.71x
20480 | 48.11 | 88.42 | 1.84x
24576 | 47.12 | 89.07 | 1.89x
27280 | 49.05 | 50.31 | 1.03x
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK][CK Tile] Grouped Convolution backward weight profiler
flush cache (#5454)
## Motivation
Flush cache to get more stable results during profiling old ck and ck
tile.
## Technical Details
Flush cache before each kernel call and one more first run.
## Test Plan
test_grouped_conv_bwd_weight_tile
## Test Result
pass
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-966
[CK] fix moe memset size which is bigger than alloc
## Motivation
Fix an out-of-bounds hipMemsetAsync in DeviceMoeGemmBlockScale that
crashes split-K MOE GEMM with "HIP runtime error: invalid argument".
When KBatch > 1, the invoker zeroes the output buffer using arg.M *
arg.N as the byte count. However, arg.M is the padded sorted-token-id
length from MOE routing, which can be much larger than the actual output
allocation (NumTokens * TopK * N). This causes hipMemsetAsync to write
beyond the buffer, and the silently-swallowed HIP error propagates to
the subsequent kernel launch via hipGetLastError().
This patch replaces arg.M with arg.NumTokens * arg.TopK so the memset
matches the actual output size.
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK Tile] Eight Waves pipeline GEMM
## Motivation
Eight waves pipeline was added for ABQuant. The goal of this PR is to
enable it also for GEMM
## Technical Details
Summary:
- Block:
- Create block struct for GEMM using eight warps specific distribution
encodings
- Use this block struct in ABQuant for encodings
- Pipeline:
- Create impl pipeline for eight waves which can be used by GEMM and
ABQuant as base (and for AQuant and BQuant in the future)
- Create eight waves pipeline for GEMM (this can not be easily
integrated in the existing async pipeline)
- Pipeline policy:
- Extract GEMM specific parts in the ABQuant policy to define GEMM
policy (then ABQuant use it as base and add Quant specific methods)
- Minor: naming was inconsistent between warp/wave, everything is now
referred to as eight waves
So overall we have:
- block struct directly used by GEMM -> ABQuant derived struct to
implement operator
- Impl base pipeline with general implementation -> GEMM and ABQuant
pipelines use it to avoid code duplication but still define their own
pipelines
- pipeline policy struct directly used by GEMM -> ABQuant derived policy
struct for Quant specific parts
## Test Plan
Added new tests for GEMM pipeline:
`test_ck_tile_gemm_pipeline_comp_async_eight_waves` (only gfx950
supports it).
Note: K padding test is disabled for this pipeline because it's not
implemented yet
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
